Satya Nadella argued in an October 10 blog post that controls on AI models must come from outside the models themselves. Project Perception is the AI security system Microsoft unveiled in July, according to GeekWire, and the new features sit in Microsoft 365 E7 and Project Perception.

On October 19, Satya Nadella posted that he recently wrote about treating models as insider risks, with security controls sitting outside the model itself, and that Microsoft is now putting that principle into practice by expanding Project Perception with new capabilities to continuously assess agent risk, adapt permissions and contain threats.

In a Microsoft Security Blog post the same day, Nadim Abdo, a corporate vice president for identity and network access, wrote that agents can make mistakes, be manipulated or be compromised, and that a model should not be its own security authority. He described four continuous jobs: know which agents exist and who owns them, contain what each can do before it acts, protect the data and tools in its path, and adapt protection as risk changes.

GeekWire reported that each agent gets a continuously updated risk assessment based on signals such as what data it can reach and whether its actions depart from its task. When risk rises, the system can remove the agent's ability to make changes while still letting it read, require a person to approve a high stakes step, or shut it down. Microsoft is also raising the number of agent types it can detect on employee computers to 60 from 25, including Claude Code, Codex and GitHub Copilot, and adding Guardian Agents that investigate why another agent's risk changed. The Next Web said a feature called Dynamic Permissions starts an agent with no access of its own and grants only what one task needs, for as long as it lasts.

GeekWire said the features are in private preview, with broader availability planned for Microsoft's Ignite conference on November 17 to 19.

Watch whether broader availability arrives at Ignite in November as GeekWire reported, and how the risk scores hold up against real agents such as Claude Code and Codex running on employee machines. GeekWire framed this as competition with Google, Amazon Web Services, CrowdStrike and Okta to become the system businesses use to control their agents, so expect rival announcements. The key open question is how often automatic permission cuts interrupt legitimate work.

This is a report based on unnamed sources, and OpenAI and SoftBank declined to comment to Reuters. Bloomberg reported earlier in October that OpenAI was seeking at least $30 billion at about $1.4 trillion, so the new detail is the faster timetable ahead of Anthropic's listing.

On October 19, Reuters posted an exclusive saying OpenAI seeks to accelerate fundraising ahead of Anthropic's IPO. The Reuters story, citing two people familiar with the matter, says OpenAI is racing to secure investor commitments before Anthropic's expected public debut in mid November, in an unusually fast round that could value it at over $1.4 trillion. One person said SoftBank, Abu Dhabi based MGX and other existing backers are in talks to join the round of about $30 billion.

According to Reuters, OpenAI met long only funds and sovereign wealth funds last week while Anthropic hosted its own investor day, and some investors see the push as an effort to lock in money that might otherwise go to Anthropic's IPO. One person said OpenAI is telling investors growth sped up after GPT 6 Astra, that it reached $50 billion in annualized revenue and expects $70 billion by year end. Reuters said OpenAI wants to finish before Anthropic's disclosures invite comparisons, and that the money would also fund infrastructure commitments and support its balance sheet if its planned 2027 listing slips.

The news wire account Walter Bloomberg summarized the report earlier in the evening. Bloomberg reported on October 5 that UAE funds including MGX discussed investing as much as $10 billion and that BlackRock was also in talks, according to Investing.com.

Watch whether OpenAI announces a closed round and its final valuation before Anthropic lists in mid November, and which investors join. Anthropic's IPO filing will give investors their first detailed look at a frontier lab's finances, which Reuters said OpenAI wants to get ahead of. Also watch whether the gap between OpenAI's net revenue accounting and Anthropic's gross figures becomes a public point of comparison.

Bloomberg reports this first hand, and Reuters and the Federal Register notice confirm it, so it is not just a rumour. Polymarket posted it earlier as a secondhand reported headline. The order covers civilian agencies only, and BBC reported that Defense Department agencies are excluded.

On October 19, Bloomberg posted that the Trump administration barred civilian US government agencies from using products or services from DeepSeek, whose more affordable open weight models pose an increasing challenge to Silicon Valley developers. Polymarket had posted earlier that morning that the administration reportedly ordered DeepSeek removed from federal agency systems.

Reuters reported that the Department of Homeland Security is ordering the removal of DeepSeek products and third party services using the model from civilian agency IT systems, according to a notice published Monday, and is banning agencies from acquiring them. DHS said DeepSeek models pose multiple threat vectors, that use may result in the transfer of sensitive data to China, that it could produce actionable intelligence on US persons or agencies, and that the models themselves may contain malware or other security vulnerabilities. Reuters said a July DHS memo, also made public Monday, told agencies to monitor traffic to and from government websites for DeepSeek risks.

The BBC reported that the ban covers civilian agencies such as Treasury, State and Justice but excludes Defense Department agencies, that it took effect from the end of July, and that it extends to technology built with the aid of DeepSeek's models. The Federal Register lists the DHS order for publication on October 20.

Reuters noted that Commerce, the Navy and several states had already banned DeepSeek on government devices.

Watch how agencies and contractors identify products built on DeepSeek models, since the ban reaches third party services and many tools quietly use open weights. Also watch whether DeepSeek or Beijing responds, whether the Defense Department issues its own order, and whether the rule spreads to other Chinese open models such as Qwen or Kimi. The same day, The Information reported DeepSeek is working with Huawei on chips, a reminder that the split runs through hardware as well as software.

Bloomberg reports this first hand, citing people familiar with the negotiations, and Reuters said it could not independently verify it. The Kobeissi Letter, a markets commentary account, posted it first as a secondhand headline. Neither company had commented, and the deal is not final.

On October 19, Bloomberg posted that Nscale, the AI cloud computing firm, is in talks to sell computing power to Anthropic in a deal that could be worth about $45 billion, according to people familiar with the negotiations. The Kobeissi Letter had posted the news minutes earlier, citing Bloomberg, adding that a deal could be finalized as soon as the coming days and would bolster Nscale's IPO prospects.

Reuters' summary of the Bloomberg report says Nscale is seeking the deal for a second facility at its Monarch data center campus in West Virginia. In August, Anthropic agreed to spend $45 billion renting computing power from that campus. The report says a second Monarch building would deliver 460 megawatts using Nvidia's Vera Rubin chips, that Anthropic has a right of first refusal on an expected third building, and that discussions are ongoing and details could change.

Reuters also noted that Bloomberg reported a group of investors had urged Anthropic to lower its IPO valuation target after market setbacks for AI companies, and that rival CoreWeave's shares fell last week over its debt load. Reuters said Nscale's total contracted value has grown from $100 million to more than $103 billion in two and a half years.

The talks follow Microsoft's exit from a flagship Nscale site in the UK over power connection delays, covered in the October 18 deck.

Watch whether the deal is signed in the coming days, as Bloomberg's sources suggested, and whether Anthropic or Nscale confirm the size and terms. If it closes, Anthropic's commitment to Nscale would roughly double, which matters for Nscale's IPO after losing the Microsoft site. Also watch whether investors keep pressing Anthropic to lower its IPO valuation, and whether power delays like those that hit Nscale in the UK spread to West Virginia.

Clash Report, a news aggregation account, posted a short clip from the Invest in Germany Summit in Berlin. The October 15 deck covered a Handelsblatt report, relayed by Walter Bloomberg, that Germany was preparing this partnership, and the summit is where it was announced.

On October 19, Clash Report posted a clip of Nvidia chief executive Jensen Huang telling Chancellor Friedrich Merz that he had heard Merz wants to find a balance between innovation and regulation, and that there is nothing to regulate if the industry does not thrive.

The exchange came at the Invest in Germany Summit in Berlin. According to Nvidia's blog, Merz and two federal ministries signed a joint declaration of intent with Nvidia earlier that day to work on an initiative advancing AI models for physical AI, and to expand Nvidia's work with Germany's Industry 4.0 Platform of government, industry and research groups.

On stage, Merz urged Germany's small and medium sized companies, the Mittelstand, to open your doors for AI to make use of your data. Huang told the audience their industrial knowledge cannot be replaced and that AI can only augment it.

CNBC reported that Nvidia named four German startups as investment targets, Agile Robots, Avelios Medical, Mirelo and Reflex Aerospace, as part of a nearly $10 billion commitment to Europe's AI ecosystem, without disclosing amounts, and that Siemens announced a wider partnership with Nvidia on autonomous factories. Reuters reported that Merz is under pressure to revive growth as support rises for the far right Alternative for Germany party.

The Handelsblatt report covered on October 15 had said hundreds of millions of euros in public money would back the partnership. The public coverage of the declaration did not give a figure.

Watch for the size of Germany's public funding and of Nvidia's investments in the four startups, neither of which has been disclosed. Also watch how the partnership fits with the EU's AI Act and its rules for AI systems, since Huang's remark pushes against tighter regulation. Merz said he wants regular meetings with Huang to check progress, which will be the next public signal of whether the initiative is moving.

The benchmark figures are self reported by OpenAI on two evaluations it built itself, MultiLangBench and a COBOL migration test. OpenAI calls these languages low resource because there is much less public code and documentation for models to learn from than for Python or JavaScript.

On October 19, OpenAI posted that its latest model advances bring stronger capabilities for building and migrating software in low resource programming languages such as COBOL, Fortran and Visual Basic, which it said power software that banks, governments and businesses depend on. It asked teams working on these systems which languages and workflows it should tackle next.

In its accompanying post, OpenAI described two evaluations. MultiLangBench asks a model to build a SQLite compatible database engine from scratch in one of 20 languages, checked against tests adapted from SQLite's own suite. OpenAI said GPT 6 Astra reached an average pass rate of 87.4 percent across the 20 languages at maximum reasoning effort, up from 26.8 percent for GPT 5.6 Sol, with COBOL rising from 10.9 to 72.1 percent, Fortran from 21.4 to 83.5 percent, OCaml from 25.6 to 85.8 percent and Visual Basic from 27.5 to 89.4 percent. COBOL remained the hardest language.

The second evaluation tests migrating existing COBOL programs to another language while keeping their behavior. OpenAI said GPT 6.1 Sol scored 86.6 percent at low reasoning effort, against 5.4 percent for GPT 5.6 Sol, at 39 percent lower cost, and 97.9 percent at higher effort, which it said nearly saturates the test. It said the work was guided by developers and enterprise customers who told it coding models often struggle with these languages.

Crypto Briefing noted that Anthropic and IBM are also pursuing COBOL modernization.

These are OpenAI's own evaluations, so the key test is whether independent teams and enterprise customers see similar results on real legacy systems, where undocumented business rules and dialects are common. Watch whether OpenAI publishes the MultiLangBench setup so others can rerun it, which coverage of the announcement says it plans to do, and whether banks or government agencies announce migrations using these models. Also watch how Anthropic and IBM respond in the same market.

The generalization and price claims are self reported by Dream Labs, which calls Dream 1 a research beta not validated for safety critical use. Dream Labs says it raised $90 million in seed funding, and Forbes reported the team are former Nvidia researchers.

On October 19, Dream Labs posted that it is coming out of stealth with Dream 1, a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3 an hour. It said its Dream API is live for everyone today.

The company's research post calls Dream 1 a World Action Model, which learns to predict video and actions together. Its argument is that through the head camera every robot is just shapes moving through a scene, so if the model can predict what a robot will look like while completing a task, it can produce the actions to get there. Besides joint angles, it can output where each gripper should be and whether it is open, a format shared by any two armed robot with parallel grippers, and each robot's own inverse kinematics solver turns those poses into joint commands.

Dream Labs says no data from the specific robots was used for its zero shot tests, that it even controlled the original PR2 research robot released 18 years ago, and that performance varies by robot and by the quality of its solver. It says developers can send camera images, the robot's state and a task to the API and get actions back, and can fine tune the model on their own robot with demonstrations.

Watch for independent users reporting how well Dream 1 works on their own robots, since the zero shot results come from Dream Labs' own tests and the company says performance varies. The limit to two armed robots with parallel grippers is a meaningful constraint, so watch whether it extends to hands or other bodies. Also watch how its price compares with robot models from Google, Nvidia and Physical Intelligence.

Ant Ling is the AI model arm of Ant Group's AGI initiative, publishing under inclusionAI. The benchmark charts are self reported by Ant Ling, though two of them are drawn from the Artificial Analysis leaderboards.

On October 19, Ant Ling posted that it is open sourcing five Ming 3.0 models across three speech families. It listed a 0.5 billion parameter text to speech model for voice design and paralinguistic cues, a Tiny text to speech model with 7.9 billion total and 1.3 billion active parameters for long form speech, a Flash audio model with 126 billion total and 6.5 billion active parameters for understanding and tool use, a Tiny audio model for efficient dialogue, and a Tiny speech recognition model with 8.6 billion total and 2 billion active parameters for dialects and context.

The overview slide says the text to speech models handle voice design, voice cloning and natural language control of delivery across seven languages and two dialects, while the speech recognition model handles dialect recognition, hotword lists and conversation context.

In Ant Ling's charts, the text to speech model leads an Artificial Analysis pronunciation robustness chart at 93.2 percent, ranks first on an open weights leaderboard using cloned voices and third on one using provider voices. Its own charts show the Flash audio model ahead of Seed 2.0 Lite and Mimo V2.6 Pro on several audio benchmarks but behind Qwen3.8 Omni Flash or Gemini 3.8 Flash on some.

OrcaRouter, which tracks model releases, noted that some weights had been on Hugging Face since September with little attention before the code and this announcement arrived.

Watch whether developers adopt the Ming 3.0 models, measured by Hugging Face downloads and third party integrations, since OrcaRouter found early downloads were low. Independent tests of voice cloning quality and dialect recognition will show whether the leaderboard positions hold up. The open weights also raise the usual question of misuse for voice cloning, so watch for any safeguards or license terms Ant Group adds as the models spread.

This is a secondhand headline from Walter Bloomberg, a news headline account, relaying a paywalled report from The Information that cites unnamed sources. Neither DeepSeek nor Huawei is quoted confirming the chip design work, though DeepSeek announced a software partnership with Huawei in September.

On October 19, the headline account Walter Bloomberg posted that DeepSeek and Huawei are collaborating on new AI chip designs to challenge Nvidia, citing The Information.

The Information's article, by Qianer Liu and Juro Osawa, is paywalled. Techmeme's summary of it says, citing sources, that the two companies are collaborating on future AI chip designs, and that DeepSeek plans to start using Huawei's Ascend 950DT for pretraining as soon as the end of 2026.

The report builds on a relationship that is already public. On September 30, Reuters reported that DeepSeek said it had partnered with Huawei to develop programming tools optimized for Huawei's Ascend chips, that it was open sourcing programming infrastructure for the Ascend platform, and that the two had jointly advanced a supernode system based on 128 Ascend 950 chips.

According to an earlier Reuters summary of Huawei's roadmap, the Ascend 950DT is the version of the 950 meant for model training and the decode stage of inference, and Huawei plans further Ascend 960 and 970 series in later years.

The post itself adds no detail beyond the headline, and what the collaboration on chip designs involves is not described in the public summaries.

Watch whether DeepSeek or Huawei confirm the chip design work, and whether DeepSeek's next model is actually pretrained on Ascend 950DT chips, which would be the clearest sign that Chinese labs can train frontier models without Nvidia. Also watch Huawei's production volumes for the 950 series and any new US export rules aimed at the partnership. Investors will read any confirmation as pressure on Nvidia's China business.

The benchmark and speed figures are self reported by fal. The Elo ranking comes from Artificial Analysis Video Arena, which fal describes as preliminary, and the latency numbers come from fal's own tests. fal is a generative media cloud, and Black Forest Labs built the FLUX 3 base model.

On October 19, fal posted that it is introducing FLUX 3 Max, the latest post trained video model from fal Research in partnership with Black Forest Labs. It said the model is tied for first in image to video quality on independent third party benchmarks, generates a 10 second video at 720 or 768p in about 12 seconds, and is 75 percent off for the next three weeks, making it what fal calls the most affordable frontier video model at its quality level.

The chart in the post gives FLUX 3 Max an Elo rating of 1139 on image to video, matching Seedance 2.5 at 1139 and ahead of Gemini Omni Flash 1.1 at 1134, H3 Max at 1124, Wan 3.0 at 1127, FLUX 3 at 1089 and Grok Imagine Video 1.5 at 1068. It lists an average latency of 12.45 seconds for a 10 second video, against 277.80 seconds for Seedance 2.5 and 126.44 seconds for the original FLUX 3.

fal's press release says the 1139 rating came from 4,868 votes in preliminary human preference tests, with confidence intervals overlapping the top models, and that the model was improved through post training, reinforcement learning and inference optimization. Black Forest Labs chief executive Robin Rombach said FLUX 3 was built as a foundation for others to build on. fal's page lists a launch price of $0.0425 per second at 720p until November 7, against a list price of $0.17.

Watch whether FLUX 3 Max holds its tied first place as the Artificial Analysis arena collects more votes, since fal calls the current ranking preliminary. Developers will also test whether the speed holds up under real traffic and what happens to demand when the discount ends on November 7. More post trained versions of open foundation models, built by inference companies rather than the original labs, could follow if this one catches on.

The reliability figures are self reported by Meta. Prometheus is Meta's supercluster in New Albany, Ohio, and Meta released Muse Wonder on October 12, saying it was trained there.

On October 19, AI at Meta posted that it trained its Muse foundation models, including Muse Wonder, at Prometheus, its gigawatt scale supercluster in New Albany, Ohio.

Meta wrote that across a supercluster with hundreds of thousands of GPUs, even one underperforming host can slow an entire training run, so it engineered Prometheus for resilience at scale. It said the cluster achieved over 85 percent goodput on pretraining and over 90 percent on reinforcement learning, meaning the share of time spent on useful training rather than lost to failures and restarts, and a mean time between interruptions of about 47,000 GPU hours while training Muse Wonder. The post linked to a technical write up on the systems behind Prometheus.

The 24 second video opens on an aerial shot of the campus with the title Inside Prometheus, then moves through rows of server racks, cabling and networking hardware before ending on the Meta logo.

StorageReview, covering the write up, described Prometheus as scaling toward 1 million GPUs on Nvidia GB200 and GB300 NVL72 systems.

Meta's Muse Wonder announcement on October 12 said the model family was trained at Prometheus, covered in that day's deck.

Watch whether other labs publish comparable goodput and interruption figures, which would let outsiders compare how efficiently the largest clusters run. Meta's next test is whether Prometheus keeps these numbers as it scales toward a million GPUs, and whether the next Muse models arrive faster as a result. Also watch local debates over power use at gigawatt sites like New Albany, which Meta addresses in its own page on the community.

The speed, cost and accuracy figures are self reported by Skyfall AI and the model is in research preview, not generally available. Sam Pasupalak is Skyfall's cofounder and chief executive and previously led Maluuba, which Microsoft acquired.

On October 19, Sam Pasupalak posted an introduction to SKY 001, which he called the first engineering world model. He said that on CAD and BIM tasks it beats GPT 6 Astra and every frontier model the team tested, running 51 to 125 times faster and 150 to 210 times cheaper with a higher intersection over union score, a measure of how closely generated shapes match targets. He said the model took 18 months to build and that the first AI general engineer will be the backbone for autonomous manufacturing.

Skyfall's press release says SKY 001 is in research preview. It describes a hierarchical visual world model that, from text or a 2D sketch, plans the full sequence of CAD steps, breaking a design into parts down to individual sketch strokes and extrusions and then reassembling them into a 3D object that opens in tools such as Autodesk Fusion. The company says this differs from generative models that produce only surface geometry without the feature trees engineers need.

The product page says SKY 001 does not do next token prediction like a language model, and pitches future uses from circuit board and aerospace design to protein folding.

In the release, Pasupalak said replacing brute force LLM calls with long term planning world models will make token costs plummet.

Watch whether Skyfall publishes a technical report, benchmark details or open access so others can check the speed, cost and accuracy comparisons against GPT 6 Astra. The research preview label means real users have not yet tested it at scale. The useful signals are design partners in manufacturing or architecture using it on production parts, and whether larger labs answer with their own CAD capable models.

This is a secondhand post from Ed Zitron, a critic of the AI industry, about reporting by Business Insider's Eugene Kim. The article relies on internal documents, a recording and unnamed people, and Amazon disputed parts of it.

On October 19, Ed Zitron posted that Business Insider confirms Amazon has warehouses of GPUs that can't be installed, and is buying them so far in advance that a supply chain team said they didn't have a clear forecasted path to installing them.

Business Insider reported that AWS chief financial officer John Felton joked at a September staff meeting that infrastructure executive Prasad Kalyanaraman had taken the company free cash flow negative, and that Kalyanaraman said the AI business is hugely free cash flow negative and would likely stay negative for the next couple of years. Amazon is spending $220 billion this year on AI infrastructure, the article said.

According to internal emails it reviewed, Amazon set up an Early Rack Receipt process for Nvidia GB300 racks, taking ownership of them earlier rather than when they reach a data center, and one email projected about $5 billion of racks in August and roughly $3 billion a month after that. People familiar with the effort said some sites had space but not enough power, so racks were staged in warehouses. One email said at least one supply chain team lacked a clear forecasted path to install them.

An Amazon spokesperson called staging hardware near sites best business practice and disputed that the company lacks a forecasted path. Zitron's framing that the GPUs can't be installed goes further than the article, which describes delays.

Later that evening, Zitron posted his opinion that NVIDIA appears to be channel stuffing, forcing Amazon and others to take early delivery of GPUs. He gave no evidence beyond the article, which does not say NVIDIA required this.

Watch Amazon's next earnings report for free cash flow, capital spending and any comment on power constraints. Business Insider said AWS still expects demand to exceed capacity in 2027, so watch whether delays in power and labor slow AWS growth. Also watch whether Microsoft, Google and Meta report similar staging of equipment, since the article said the problem is not unique to Amazon.

Elice is a South Korean company that calls itself a neocloud, meaning it rents out AI computing capacity. Zerohedge, a markets commentary site with a strongly skeptical view of the AI buildout, reported the debut and also posted a table of other data center listings that were delayed, cut or withdrawn since September.

On the evening of October 19, US Eastern time, as trading opened in Seoul, zerohedge posted that Korea's biggest IPO of the year had cratered 32% on its debut.

According to zerohedge, citing Bloomberg, data center infrastructure company Elice raised 201 billion won, about $151 million, on the Kosdaq, South Korea's tech focused exchange. It priced at 90,500 won, the top of its marketed range, for a market value of just over 1 trillion won, about $750 million. Shares briefly opened around 78,000 won, still about 14% below the offer price, then fell as much as 32% to 61,400 won.

Zerohedge then posted a table it titled The Casualty List, crediting itself, Bloomberg, The New York Times, Reuters, The Wall Street Journal and Light Reading. It lists a delayed SB Energy IPO for SoftBank, a Firmus IPO that was cut and then withdrawn, a deferred Vitro REIT listing in the Philippines, a postponed AirTrunk REIT in Singapore, Nscale's lower valuation talks after Microsoft exited a UK deal, Elice, and DayOne as a possible next listing. The table is zerohedge's compilation, and this deck has not checked each entry.

In a separate post, zerohedge mocked reports that the offering had been heavily oversubscribed. Strong demand for an offering at pricing does not guarantee that shares hold up once trading opens.

Watch how Elice trades in the days after its debut, and whether it recovers to the offer price. The next test named in zerohedge's table is DayOne, which it says could market a US listing of up to $5 billion as soon as October 21. Nscale's planned IPO and its talks with Anthropic over more computing capacity are another signal of how public investors value AI infrastructure companies.

Hyperion is Meta's large AI data center campus in Louisiana, financed through a separate vehicle that sold about $27 billion of bonds in October 2025, as covered in this deck on October 10. Zerohedge is a markets commentary site, and the chart it posted is unsourced, though it appears to show market price data.

On October 19, zerohedge posted that the world's largest corporate bond, RPLDCI of 2049, the $27 billion bond funding Meta's Hyperion data center, hit new lifetime lows.

The chart it attached is labeled RPLDCI Sr Secured 6.581% due 2049, Project Beignet. It shows the bond's price starting near 110, meaning 110 cents per dollar of face value, in October 2025, drifting to around 102 by June 2026, then falling to about 91.5 in October 2026. Zerohedge did not name a data source.

A bond's price moves opposite to its yield, so a lower price means buyers now demand a higher return to hold it. The post does not say why the price fell. A general rise in long term interest rates would also push down the price of a bond that does not mature until 2049, so the move is not necessarily a judgment on Hyperion itself.

This deck reported on October 10 that the Hyperion financing is held off Meta's balance sheet in a separate vehicle, and that the Financial Times has reported tech groups have shifted more than $120 billion of AI data center debt off their balance sheets. Zerohedge's post comes alongside its reporting the same day on troubled data center listings, part of its broader skeptical coverage of AI infrastructure financing.

Watch whether the Hyperion bond keeps falling, and whether other bonds and loans tied to AI data centers weaken with it, which would make new financing more expensive. Meta's next earnings report may show whether it plans to fund more capacity the same way. Also watch for any commentary from rating agencies on these data center financing vehicles, and for moves in long term Treasury yields that could explain part of the drop.

This is a first hand teaser from MiniMax's official account, posted with a short video. A free anonymous model called Space Bunny Alpha appeared on OpenRouter in September, and Chinese outlet Yicai reported that the post was seen as MiniMax claiming it as an early version of M3.1 Flash.

On October 19, MiniMax posted: Welcome back. Space Bunny. MiniMax M3.1 Flash is coming. Its 25 second video shows a carrot built like a rocket lifting off from a launch tower, with Space Bunny written on its side, climbing through clouds into orbit and splitting apart above the Earth before the MiniMax M3.1 flash logo appears.

The post quoted Ryan Lee, MiniMax's head of developer relations, who wrote: Dear friends, we are back, with an image of the same carrot rocket over the Earth.

Space Bunny Alpha appeared on OpenRouter on September 23 under a provider named Stealth, with a 1 million token context window, according to CodersEra, which argued on October 5 that its settings matched a MiniMax preview model. Yicai reported that within three days of launch Space Bunny led daily token usage on OpenRouter and OpenCode, alongside DeepSeek V4.1 Flash and GLM 5.3 Flash, and that it is an early version of M3.1 Flash.

MiniMax's API documentation lists MiniMax M3.1 Flash as its latest model, with a 1 million token context window, text, image and video input, and a setting to tune how much it thinks. The post calls the model coming, and MiniMax has not published benchmark results or pricing for it.

Watch for MiniMax's formal release, including pricing, benchmark results and whether the weights are published.

Also watch how M3.1 Flash compares with DeepSeek V4.1 Flash and other small coding models in independent tests. Launching under an anonymous name first is becoming a common way for labs to test demand, and Space Bunny's usage gives MiniMax a head start with developers who already tried it.

The benchmark results are self reported by Exa, which ran all providers in its own harness with a GPT 6 Luna agent, according to the table's notes. Exa sells a search API built for AI agents and said earlier this month it passed $100 million in annualized revenue.

On October 19, Exa posted that it is introducing Exa 3.0, its largest ever improvement to search quality, saying Exa is now first on every benchmark it tested, from public evals to verticals like coding and GTM intelligence, and that this is the result of multiple research breakthroughs.

The attached table compares Exa 3.0 with Perplexity, Brave and Parallel on eight benchmarks. It shows Exa ahead on each, though some margins are small, such as 91.8 percent against Perplexity's 91.5 percent on BrowseComp and 88.2 percent against 87.6 percent on DSQA. Larger gaps appear on ATLAS, at 42.0 percent against 37.2 percent, and WebCode, at 75.5 percent against 69.9 percent. The notes say each provider's highest scoring single step search endpoint was used, all running in the same agent harness evaluated in October 2026.

Exa's blog says the gains come from a larger embedding model trained with more compute and data than any before, Dynamic Highlights, now generally available, which it says uses 95 percent fewer tokens than full page text, and billions of new documents such as ecommerce listings, local businesses and flights. It also launched Exa Lite at $4 per 1,000 requests, with Exa Auto at $7 per 1,000. Startup Fortune noted that outside comparisons show smaller leads than Exa's own tests.

Watch whether independent evaluations confirm Exa's lead, since several of its margins over Perplexity are under one percentage point. Exa plans a live Q and A on October 21, which may give more detail on its methods. Also watch how Perplexity, Brave and Parallel respond on price, since the cheaper Lite tier targets developers running many agent searches, and whether agent builders switch providers based on these results.

This is a technical report on a model Meta already released, not a new model. Meta introduced Muse Glimmer with open weights under Apache 2.0 in August, and the benchmark comparisons are Meta's own.

On October 19, AI at Meta posted that it published a technical report on Muse Glimmer, its 30B open weight model built for agentic tasks on consumer hardware. It said the model runs at practical speeds on a Mac or PC with a single consumer GPU at 4 bit precision, and leads comparably sized open models on tool use and multi step planning. It said the report covers the architecture, distillation recipe and quantization.

The report's abstract describes a 30 billion parameter language model with a perception encoder that takes text and images and outputs text, trained on data in over a hundred languages. Meta claims state of the art results for its size on agentic benchmarks against Gemma4 31B and Qwen3.6 27B.

Meta's August launch post said that at full precision the model would need over 55 GB of memory, and that quantization to about 4 bits shrinks the language model to under 20 GB, leaving room for memory, the vision encoder and a speculative decoding drafter within 24 or 32 GB. The Hugging Face model card says it was distilled from Muse Spark and lists a 131,072 token context window.

The report adds detail on how the model was built rather than new capabilities, so its main value is for researchers and developers running local agents.

Watch whether outside developers reproduce Meta's agentic benchmark results on consumer hardware, and how Muse Glimmer compares with new small models from Google and Alibaba's Qwen team. The distillation recipe in the report could be reused by others to build small agent models. Also watch whether Meta ships Muse Glimmer inside its own apps for on device features, and whether it releases a newer local model.

This is first hand from Greptile, and its usage figures are self reported. Greptile makes AI agents that review and test pull requests, and TechCrunch reported in 2025 that Benchmark was in talks to lead its Series A at a $180 million valuation.

On October 19, Greptile posted that it has raised a $50 million Series B led by Menlo Ventures, valuing the company at $750 million. It said it reviews and tests more than 1 percent of all pull requests opened on GitHub every day, and that thousands of engineering teams including Nvidia, Coinbase, Decagon and Vanta use it as their independent and central validation layer. It said it has fewer than 50 people and is hiring.

Greptile's blog post says Jump Trading, Wing, Operator Collective, Benchmark, YC, Browder Capital and Initialized also took part, and that Deedy Das is joining its board.

The launch video is a black and white parody of a retro office, presenting Greptile as the world's premier code validation facility. Its subtitles say that before code ships it comes there, that experts read it line by line, that Greptile does not stop at reading the code but runs it in a simulator and tests it, and that a change can get a confidence score of three out of five. It ends with the tagline from code review to code validation.

In 2025, TechCrunch reported that Benchmark was in talks to lead a Series A valuing Greptile at $180 million, citing sources, which would make the new valuation roughly four times higher.

Watch whether Greptile's claim of reviewing more than 1 percent of GitHub pull requests is borne out by independent data, and how it competes with Graphite, CodeRabbit and the review features built into Codex, Claude Code and GitHub Copilot. Its shift from reviewing to running and testing code is the bet to track. Also watch hiring, since a team under 50 people will need to grow to serve thousands of customers.

The figures are self reported by Figure's founder and chief executive, and the post gives no independent record of the run. In May, Adcock said a team of F.03 robots sorted packages for 200 hours, swapping in and out to recharge, according to Sherwood News.

On October 19, Brett Adcock posted that Figure's F.03 humanoid had just crossed 9 straight days walking on a treadmill, covering 348 miles and 1.2 million steps. He said the run started as a side project to stress the robot for any failures, adding: Let's see how long it can go.

The frames of the 6 and a half minute video show one robot walking steadily on a treadmill inside a glass enclosure in an office, with two charging units on the floor next to it and people occasionally passing in the background. The post does not say whether the same robot walked the entire time, whether it stopped to charge, or whether any parts were replaced, so the claim of 9 straight days cannot be checked from the video alone.

It is Figure's second public endurance test this year. In May, Adcock said a team of F.03 robots ran a package sorting line for 200 hours without a failure, after what began as an 8 hour challenge against a human intern. Sherwood News reported that in that test several robots took turns, swapping out every three to four hours when a battery ran low.

Adcock gave no end date for the treadmill run.

Watch for Figure to publish how the run was measured, including charging breaks, maintenance and whether one robot did all the walking, which would make the numbers comparable with rivals. A walking test on a treadmill is far easier than useful work, so the more important signal is whether Figure shows similar uptime on real tasks for paying customers. Also watch how long the run lasts and whether it ends with a failure or a planned stop.

This is a secondhand post: Polymarket, a prediction market, summarized a Financial Times piece without linking it. The piece is an opinion column by Chris Miller, author of the book Chip War, published October 19.

On October 19, Polymarket posted that AI data centers could soon require blast resistant concrete, anti drone defenses and missile protection as attacks on computing infrastructure escalate worldwide, crediting the Financial Times.

The source is a Financial Times opinion column by Chris Miller, the historian who wrote Chip War. Its summary says AI infrastructure designs may now need to incorporate blast resistant concrete, jamming gear and missile defense systems. Polymarket's wording that data centers will require these measures states the column's argument more firmly than a prediction.

The column points to recent attacks. The Financial Times separately reported on October 14 that Ukrainian strikes on data centers run by Yandex disrupted services in Russia, from email and taxi apps to electronic key fobs for Moscow apartment buildings.

In September, Reuters reported that the United Arab Emirates was revising a planned 5 gigawatt AI campus after Iranian missile and drone attacks on Gulf states, which Reuters said damaged two Amazon Web Services data centers in the UAE and one in Bahrain in March. People familiar with the plans told Reuters that officials were considering spreading the project across several sites, building some facilities underground, and adding air defenses and blast resistant concrete. G42 told Reuters the work was progressing as planned.

Watch whether data center developers and their insurers start to price in physical security measures such as hardened buildings and air defenses, and whether governments begin to treat AI data centers formally as critical infrastructure. The UAE's redesign of its AI campus will be an early test of what such protection costs. Also watch for further strikes on data centers in active conflicts, which would make the column's argument harder to dismiss.

Tibo leads Codex at OpenAI, and this is a first hand statement from him. OpenCode, an open source coding agent, disclosed a breach tied to a distillation attack on October 17, and its founder dax said attackers kept hitting its systems for the next two days.

Shortly after midnight on October 19, Tibo posted that his team worked with OpenCode to get its users who are also OpenAI API customers back in business, and to refund any credits that would have been drawn in the attack.

He added that people can request Daybreak Blue access to use OpenAI's cyber permissive setup for defense, and told them not to wait until they have to deal with an incident. We're here to help, he wrote.

OpenCode said on October 17 that a breach tied to a large scale distillation attack affected 1.4% of OpenCode Console users who had linked external model providers. Tibo's post indicates the attack drew on some of those users' OpenAI API credits. The next day, dax said OpenCode had faced attacks for 48 hours, that AI models had refused to help at points during the investigation, and that OpenCode had since been approved for Daybreak Blue.

OpenAI's help center describes Daybreak Blue as an approved tier of its Trusted Access for Cyber program that reduces refusals on supported models for verified defensive work, such as incident response and malware analysis, while other safeguards still apply.

Tibo did not say how many users or how much money the refunds cover, or how affected customers are identified.

Watch for a full incident report from OpenCode, which has not said exactly what was taken, and for whether other model providers announce similar refunds.

Also watch whether more small developer tools apply for Daybreak Blue after this incident, and whether OpenAI makes the tier easier to get. The case is an early test of how AI companies support third party tools whose users' keys are abused.

This is first hand from OpenAI's official developer account, but the post is a livestream link and the feature details come from its follow up post rather than from release notes. OpenAI did not give availability dates or plan limits for the features in the posts.

On October 19, OpenAI Developers posted a livestream titled by the line here is what we are cooking with Codex.

In a follow up post, the account described the stream. A dot, one of ChatGPT's always on agents, helps three people with a ramen recipe and keeps the prep in ChatGPT Space, while GPT Live 1 talks them through making it. The stream covers Codex cloud environments, where tasks can keep running after you close your laptop, and what is new in the Codex CLI: talking to Codex in your terminal, keeping track of your agents, and forking a conversation into a separate worktree.

The post said Matthew Berman and an account called ForwardEditor join to build an 8 bit ramen game with Ultrafast, OpenAI's fast mode for GPT 6.1 Sol, and play it on a ModRetro handheld. The stream also shows community projects, from a retro space shooter to a game where you draw your own racehorse. Earlier in the day the account asked followers what they had been cooking with Codex, and later posted a roundup of community projects.

The posts do not say which of the CLI features are newly released and which are previews.

Update: Later that afternoon, OpenAI's Codex lead Tibo quoted the livestream post with his Day 15 entry, writing only: Day 15/ We're cooked. He did not announce a reset or a new feature with it.

The community roundup post, Cool projects our community is cooking with Codex, attached a video reposted from another account, sound4movement, showing colorful animated radial diagrams. OpenAI Developers did not describe the projects in the post itself.

Watch the official Codex changelog for entries on cloud environments, voice in the CLI, agent tracking and worktree forking, which will show what actually shipped and to which plans. Developers will judge cloud tasks that keep running on how reliable they are and how much usage they consume. Also watch Tibo's daily Codex posts, which often follow these streams with resets or small updates.

Madry is an MIT professor on leave who served as OpenAI's head of preparedness before being reassigned to work on AI reasoning, according to Benzinga, and he announced he was leaving OpenAI in May. Jakub Pachocki, who posts as merettm, is OpenAI's chief scientist.

On October 19, Aleksander Madry posted that what he thought was leaving OpenAI turned out to be more of a sabbatical. He wrote that when he left a few months ago, he felt the most important thing was making AI more useful to people, but that the field has since entered a different regime of AI progress, one in which getting the safety and security challenges of increasingly capable AI right has to become a central priority. He said he will return to OpenAI, joining Jakub Pachocki to help chart a path through these challenges.

Pachocki quoted the post, writing that as AI becomes more capable he is focused on questions around the pace and implications of progress and how they should inform the strategy of OpenAI and the wider field. He said he asked Madry to come back to work with him on these topics and values his judgment.

Benzinga reported that Madry joined OpenAI in 2024 as head of preparedness, a role focused on identifying and addressing potentially catastrophic risks from advanced models, and that when he left in May he said he wanted to focus on how AI could transform the economy. Before OpenAI he directed MIT's Center for Deployable Machine Learning.

Neither post gives Madry a title or describes a specific team or project.

Watch for a title or team announcement that shows how much authority Madry will have, and whether OpenAI publishes new safety or preparedness work under him and Pachocki. His phrase about a different regime of AI progress invites questions about what OpenAI has seen internally, so watch whether the company explains it. The return also lands as OpenAI is reported to be working through safety and security concerns about its next flagship model.

This is a company announcement from Nebius, a European headquartered AI cloud provider, made with NVIDIA and the venture firm Accel. Nebius says it launched the program globally in September, and grants are awarded at its discretion.

On October 19, Nebius posted that its AI Builder Program is expanding across Europe with $10 million in total support across compute grants, capital and community programs.

The post said the program offers compute grants for selected projects in areas such as open models, coding agents, post training and evaluations, connections to venture capital firms to turn open source projects into companies, and hackathons and demo days with communities including AI Tinkerers, Tech Europe and ESRA, the European Student Robotics Association. Nebius said it would kick off the expansion this week at NVIDIA's GTC conference in Berlin, with NVIDIA, Accel and other European venture firms.

Nebius's blog post says it is looking for work created in Europe that strengthens open source AI, including open weight research, agent runtimes, inference tooling and reinforcement learning. It quotes Matt Weigand, a partner at Accel, saying the firm believes open source plays a critical role in AI. The blog says thousands applied to join within weeks of the September launch.

The program's site says basic membership is free and includes small credits from Nebius and partners such as LangSmith and Tavily, while compute grants are subject to review and availability. Nebius did not say how the $10 million is split between grants, capital and events. The video shows the program's headings and a speaker at a builder event, ending with Nebius and NVIDIA logos.

Watch which projects win compute grants and whether Accel or other firms invest in program members, which would show whether the program creates companies.

Also watch how Nebius's push competes with European efforts from other cloud providers and governments to support local AI builders.

Nebius said the expansion starts this week at GTC Berlin, so the first hackathons, demo days and grant announcements may follow soon. Details on how much of the $10 million goes to each part of the program are still to come.

The customer count and other figures are self reported by Context.dev's founder and chief executive. Context.dev was in Y Combinator's Summer 2026 batch and was called Brand.dev until it rebranded earlier this year.

On October 19, Yahia Bakour posted that his company raised $9 million to give your agents the internet, led by General Catalyst with Orange Collective and Robinhood Ventures. He said its single API lets developers search, scrape, crawl and monitor the web, is trusted by more than 800 customers, and aims to make it ridiculously easy to build with web data.

The company's blog post says it is a seed round. It says Context.dev started as Brand.dev, an API for company logos, colors and fonts, and became Context.dev this year as customers asked for products, page text and whole site crawls, and as agents needed a way to read the internet. It says teams use it to keep data on more than 20 million companies fresh and to power more than 3,000 integrations, and that it has never shipped a breaking change. New sign ups in the next 24 hours get 50,000 free credits.

Its Y Combinator launch page, from about three months ago, cited more than 400 customers and a team of four, and said Bakour previously led engineering at Sunrun, worked at Amazon, and built and sold two startups.

The launch video shows Bakour explaining the product, with diagrams of the web sources it collects from.

Watch whether Context.dev can turn its customer count into large, lasting usage against rivals that Dealroom names, such as Firecrawl, Exa and Parallel. Dealroom also notes reports that big websites are starting to block AI agents, which could make reliable access more valuable or harder to provide. Also watch how the team grows with the new funding, since it was four people at launch.

This is a self reported claim from the owner of X, and the chart shows X's own internal metric without a published definition. X publishes parts of its recommendation code on GitHub, but that code does not show the new models.

On October 19, Elon Musk posted two important changes to timeline recommendations on X: a new AI and SI ranking model that predicts probable time spent much better, and a retrieval model trained to surface posts people read all the way through. He added that user active seconds is growing exponentially and hitting all time highs almost every day.

The attached chart, titled Daily UAS, compares daily totals for 2023 through 2026. The 2026 line sits well above earlier years for most of the year and climbs past 500 billion in October, while 2023 to 2025 stay roughly between 300 billion and 400 billion. The chart does not explain how active seconds are counted, and its rise over the past few weeks is short compared with the whole series.

In a reply, Musk wrote that rewriting the X Android app was also a big deal, saying it was really bad for such a long time, which suggests the app rewrite may also have contributed to the growth.

In X's public algorithm repository, retrieval selects candidate posts and ranking orders them by a weighted mix of predicted actions that include dwell time. The new models' code and weights have not been published there, so outside researchers cannot yet check what changed.

Watch whether X publishes the new ranking and retrieval models in its open repository, and whether third party data such as app analytics firms show the same jump in usage. Optimizing for time spent and finished reading favors longer posts, so watch how creators and advertisers respond. Also watch whether the gains last once the new models are no longer novel.

The speed, price and benchmark claims are self reported by Nace AI, a startup building small decision models. Its Decision Index benchmark is published by Nace itself, and its docs list Drex Max as released on October 13.

On October 19, Nace AI posted that Drex Max is here, calling it its multimodal decision model for text, images and video. It claimed 34 ms per decision, under one second at 262K tokens, and a price of $0.06 per million input tokens, available on its site.

Unlike a chat model, a decision model writes no text. According to Nace's docs, a developer sends a state and up to 16 named questions, each a yes or no, a choice among given labels, or a score, and gets back a probability for each option in one call. Drex Max is the only Drex model that reads images, up to eight per request, and output tokens are not billed.

Nace's product page claims a score of 67.59 on its own Decision Index, 6.9 points above a rival model called Jev 1.14, across 37 public benchmarks. It also lists frontier models such as GPT 6 Sol and Claude Opus 5.5 for reference, which beat Drex Max on several general tests like HLE and MMLU Pro.

Some figures do not match across Nace's own materials. The post says 34 ms per decision, while the product page gives a median API latency of 183 ms, and the page prices Drex Max at $0.06 per thousand decisions. The post mentions video, but the docs list only image formats. The demo video shows it labeling a webcam feed and choosing clicks in web pages.

Watch for independent tests of Drex Max on computer use benchmarks such as OSWorld, since Nace runs the Decision Index itself. Also watch whether agent builders adopt decision models alongside general models to cut costs, and whether Nace clarifies its latency and price figures. Its open weights Drex 2.0 lets others test the approach on their own hardware, which could show how much of the speed comes from the model itself.

This is a promotional post from Google's Gemini account about a demo built by its own team. Wizard Chess 3D was already listed among the Gemini 4 Argon launch apps in Google AI Studio, so the game itself is not new.

On October 19, the Google Gemini account posted that its team built a playable 3D Wizard Chess game from scratch with Gemini 4 Argon, using Google Antigravity in the Gemini app for Mac. It said users can coordinate a multi agent Blender workflow to model, design and deliver fluid, real time animations, going straight from idea to play.

Google AI Studio's page for Gemini 4 Argon lists Wizard Chess 3D as one of its showcase apps, describing interactive 3D wizard chess inside a Gothic cathedral with sculpted stone pieces, animated captures and shattering stone physics, which people can try or remix.

Google's help page says Antigravity gives the Gemini app for Mac a dedicated space to build with coding agents. An agent plans and builds a project, brings in other agents when things get complex, shows a step by step plan before touching anything, and only works inside folders the user chooses. Google's guide on building 3D games says Argon can write Python scripts to model detailed objects in Blender, export them and load them into a Three.js scene.

The post did not include the prompts used, how long the build took, or how much human editing was involved.

Watch whether people outside Google can reproduce results like this from the Gemini app, which will show whether the multi agent Blender workflow is practical for everyday users or mainly a showcase. Also watch which subscription tiers get Gemini 4 Argon inside Antigravity, and how it compares with Codex and Claude Code for building games. Google may publish the full prompts or a walkthrough, which would make the demo easier to judge.

This is an unconfirmed report. Stephanie Palazzolo writes The Information's AI Agenda newsletter, and the Anthropic details rest on one person who has used the model. Anthropic, Google and OpenAI have not announced any of the three models, their names, dates or prices.

On October 19, Stephanie Palazzolo of The Information posted that this month should be a busy one for new model releases, and that the publication has the scoop on a new Anthropic model, Google's Gemini 4 Carbon and a potential new Terra model from OpenAI. The article is paywalled.

An excerpt shared by the account Chubby says Anthropic is preparing to release a new model that is better in its reasoning and intuition than the Anthropic models currently available commercially, according to a person who has used it. That person said the unreleased model is especially good at discerning a person's underlying intentions and goals even when they give it little information. Chubby added that, when asked to fix one bug, the model can reportedly identify the broader class of problems behind it and ask whether to address those too, and that the model family and release date remain unclear. Chubby's description of this as official is not accurate, since Anthropic has not confirmed it.

StreetInsider's summary of the report says Gemini 4 Carbon is an upgrade to Gemini 4 Argon, which launched last week, and that developers told The Information Google's models remain strong at visual tasks such as video analysis and website testing. It says OpenAI is preparing a new model in its Terra family, potentially for release in the coming weeks, expected to offer improved capabilities at a lower price than GPT 6.1 Sol.

TestingCatalog, which tracks AI leaks and rumors, shared more lines from the article, including that the model is better at thinking outside the box, and that it is not clear whether it will be a Fable or Opus model. Separately, the account Lumina posted that Microsoft updated 18 evaluation configs from GPT 5.6 Terra to GPT 6.1 Terra in a pull request, which it read as a sign the model is near. That is an unconfirmed leak, and OpenAI has not announced GPT 6.1 Terra.

Treat every detail here as unconfirmed until the labs announce something. The Anthropic claims come from a single tester, and the Gemini 4 Carbon and Terra details come through a secondhand summary of a paywalled article. Watch for official model cards, names and prices from Anthropic, Google and OpenAI, and for whether all three actually ship before the end of October as Palazzolo suggested. If they do, independent benchmarks will show whether the reported gains in reading intent are real.

This is a report based on unnamed people familiar with the matter. Schmidt and Anthropic declined to comment to Bloomberg, and it is unclear whether he has sold any of his stake.

On October 19, Bloomberg reporter Natasha Mascarenhas posted that one of the more interesting things about Anthropic's cap table is that until its Series C, its funding rounds were dominated by super angels and nonprofits. She shared a scoop with Tom Maloney on one of the angels that got in early, former Google CEO Eric Schmidt.

In the Bloomberg story, as republished by Business Standard, people familiar with the matter said Schmidt owns around 1 percent of Anthropic after investing tens of millions of dollars in its Series A. Bloomberg cited Pitchbook data valuing Anthropic at $461 million when he first invested in 2021. It said investors indicated last week that a fair value for the company is $1.8 trillion to $2 trillion, which would raise his net worth by about a third to nearly $80 billion, and that even at Anthropic's last valuation of $965 billion his stake would add nearly $10 billion to a fortune of $57.5 billion.

Bloomberg said Anthropic could start formally marketing its IPO as soon as the week of November 9. Reuters, summarizing the scoop, noted it had reported that some investors discussed a valuation of around $1.5 trillion at an investor meeting, below earlier expectations.

Watch Anthropic's public IPO filing, which should show major shareholders and whether early angels like Schmidt still hold their stakes. The final IPO valuation will decide whether his gain is closer to $15 billion or $20 billion. Also watch how OpenAI's race to raise money before the listing affects demand for Anthropic shares, and whether other early angels are named before the listing.

This is a secondhand post from Pablo Marti, a tech investor, sharing charts he attributes to Deutsche Bank, and the full research note is not public. The charts compare each app's first 36 days in the US, not their current usage.

On October 19, Pablo Marti posted: Deutsche Bank: Muse remains significantly ahead ChatGPT in terms of downloads and DAUs through 36 days since launch, tagging Meta's stock ticker.

The two charts are labeled US only and measured in thousands. One shows daily active users after launch, with Muse climbing to roughly 2.3 million by day 36 while ChatGPT, Sora, Gemini and Perplexity stay below about 400,000 at the same point after their own launches. The other shows cumulative downloads, with Muse at roughly 9 million by day 36, ahead of Sora at about 4.5 million and ChatGPT at about 3.4 million.

Meta's chief AI officer Alexandr Wang quoted the post, writing that this is what we in the business call framemogging, a slang boast. TechCrunch reported in September, using Apptopia data, that Muse was outpacing ChatGPT's early mobile launch.

The comparison has limits. ChatGPT's app launched in 2023 into a much smaller market for AI apps, while Muse benefits from Meta's existing apps and users, so matching launch days does not compare the products as they are today. The note's methods are not public.

Watch whether Muse keeps its daily users after the launch period, since some outside trackers have reported slowing downloads. Meta's next earnings call may give official usage numbers, which would be more reliable than bank estimates. Also watch whether ChatGPT and Gemini respond with features aimed at the same consumers, and whether Deutsche Bank's full note becomes public so its methods can be checked.

These charts are a secondhand analysis. Brandon Carl, who posts about markets and AI, built them from OpenRouter rankings data, and OpenRouter itself did not publish them. OpenRouter is a service that routes developers' requests to many AI models, so its numbers do not include traffic that goes straight to OpenAI, Anthropic or Google.

On October 19, Brandon Carl posted that OpenRouter's weekly numbers are in. He wrote that overall usage was down slightly, the first weekly decline in about a month, that Opus continues to keep Anthropic back on the map, and that Google's model releases haven't done much overall and seem to be having trouble finding their footing.

His charts label the week of October 12 with a 4.7 percent week over week drop in total tokens. For the latest week they show 79.2 trillion tokens for providers assigned to China against 17.4 trillion for American providers, 6.21 trillion for OpenAI, 6.18 trillion for Anthropic and 131 trillion in total excluding DeepSeek. Google's chart shows 0.00 trillion for the latest week, which may reflect how his data is labeled rather than true usage.

The account AllThingsVentured quoted the post with a crop of the Chinese, American and total charts, writing that US model token consumption stopped growing 8 weeks ago, that OpenAI is falling off a cliff, and, sarcastically, that it is a good thing these companies have 10x the effective compute coming online in the next 24 months. That reading is commentary, and the charts cover only OpenRouter traffic.

One weekly drop is not a trend, so watch whether the decline continues in the next OpenRouter numbers. The gap between Chinese and American models on OpenRouter likely reflects price sensitive developers, and may not match the labs' own revenue. Watch whether OpenAI's new models or Google's upcoming Gemini 4 Carbon change the picture, and treat claims about the AI compute buildout based on this data with caution.

This is a promotional launch post from Sunra, a startup that sells a creative suite and API for generative models. The no markup pricing claim is Sunra's own, based on comparisons it published.

On October 19, Sunra posted that it is launching Sunra MCP, with the most accessible pricing in the industry. It said every leading image, video and audio model is available right inside Claude, ChatGPT and other assistants, each at its maker's own API price with no markup added.

MCP, the Model Context Protocol, is an open standard that lets AI agents use outside tools. Sunra's setup page says users can add it as a connector in Claude or a custom plugin in ChatGPT, signing in through OAuth without an API key, and that Claude Code, Codex, Cursor, Antigravity and Grok Bot can set it up from one prompt. The page lists models such as GPT Image 2.5, Nano Banana Pro, Seedream 5.0 Pro, Seedance 2.5, Kling 3.0, Veo 3.1, ElevenLabs Music and Meshy for 3D.

Sunra's launch blog explains its credits and publishes tables comparing its prices with those of other platforms. Its own tables show some rivals cheaper for certain models, such as several ElevenLabs audio tools, so the claim of the most accessible pricing does not hold for every model. The attached video is an ad showing scenes with model names and starting prices.

Watch whether Sunra can keep pricing at the makers' rates as usage grows, since it earns nothing on markup and will need another source of revenue. Also watch whether Anthropic and OpenAI add similar built in media tools that make third party connectors less necessary. Independent price checks will show how often Sunra is actually cheapest, and user reviews will show whether the connectors work reliably inside each assistant.

This is self reported usage data from Theo, who runs T3 Code and T3 Chat, covering only his own product's users. He did not share the underlying numbers or how usage was counted.

On October 19, Theo posted that for the last two weeks, over 60 percent of T3 Code users who used Codex also used Claude Code. He added that the majority of Codex users are Claude Code users.

The post gives one statistic and no chart. It does not say how many people use T3 Code, how many used Codex in that period, or whether the overlap means people switch between the two within the same projects or simply tried both. Since the statistic comes from a tool where both agents can be used, its users may be more likely than average to try several, which would make the overlap higher than among developers in general.

Theo has shared his own usage data on Claude and Codex before, including a chart on October 14 of how models behave in his requests, covered in that day's deck.

The numbers land as OpenAI and Anthropic compete heavily for developers, with OpenAI's Codex team running daily usage resets and Anthropic promoting Claude Code. A large overlap would suggest that loyalty to one coding agent is weak and that developers move to whichever gives them the most usage or the best results that week.

The statistic covers only Codex users, so it says nothing about how many Claude Code users also use Codex.

Watch whether Theo or other tool makers that support several agents, such as Cursor, publish similar overlap data, which would show whether developers really mix agents. Also watch whether OpenAI or Anthropic change pricing or usage limits to win back users who split their time, since a weak lock in means each lab can lose share quickly. The reverse number, how many Claude Code users also use Codex, would complete the picture.

This is a promotional launch post from Viggle, the company behind a meme making app it says has over 50 million users. The speed, cost and benchmark claims are self reported, and the phrase world's first is Viggle's own description.

On October 19, Viggle posted that it is releasing JST 2, the world's first full stack physics native world foundation model. It wrote that many world models learn in pixel or video space, while it built the entire model stack in 3D, so users can generate a 3D scene in seconds, expand it in real time and create characters with motion built in through one architecture.

Viggle's launch page says JST, short for Joint Space and Time, represents the world as persistent atoms in 3D space, each with an explicit position and learned vectors for its attributes and dynamics. It says JST 2 generates a 3D scene in about one second on an Nvidia B200 at an inference cost below one cent, supports up to 64 input views, and outputs images or 3D Gaussian splats that can be explored in a browser. Its character page says characters can be animated without separate rigging, and its motion page says JST 2 creates full body motion from text or video.

Viggle says scene collisions and object interactions in JST 2 still rely on rule based physics, and that it is developing JST 3 to replace those rules with learned dynamics.

Watch for independent comparisons against other world models, since Viggle's speed and quality figures come from its own tests. Its own note that collisions still use rule based physics shows how far physics native still is from learned physics, so JST 3 is the step to watch. Also watch whether game and animation studios use the free trial and API in real production.

This is a secondhand post from DogeDesigner, a fan account. The page appears in Tesla's 2026 annual report and proxy statement, filed with the SEC on October 19, rather than in the letter to shareholders itself, which is signed by board chair Robyn Denholm.

On October 19, DogeDesigner posted that Tesla shared an image of Optimus in its latest letter to shareholders, with a message from Elon Musk: The most likely outcome is an age of Amazing Abundance, where anyone can have anything they can think of. This is a message of optimism and excitement about the future.

The image and quote appear on an early page of Tesla's 2026 annual report and proxy statement, which Tesla filed with the SEC that morning. The page shows the robot in the dark, with the quote attributed to Musk at the top and the Tesla wordmark at the bottom. The facing page lists V3 Optimus as readied for mass production, and the filing does not say when or where Musk first said the line.

The letter to shareholders that follows is signed by Robyn Denholm. It says Tesla began producing Cybercab, expanded its Robotaxi service, passed 10 million vehicles delivered, and built the first production lines for Optimus and more data center capacity to train its AI systems, which the filing calls super intelligence, or SI. It ends by saying Tesla is bringing SI into the physical world.

The filing asks shareholders to re elect Musk, Denholm and JB Straubel to the board at the annual meeting on December 3. Tesla's voting site uses the same theme, Building a World of Amazing Abundance.

Watch the December 3 shareholder meeting, where Musk is up for re election and is likely to give an update on Optimus production.

Also watch for evidence that V3 Optimus actually reaches mass production, and for details on what the robots will do and cost. Investors will compare the filing's claims, such as V3 Optimus being ready for mass production, with actual deliveries.

This is an opinion post from Musk, whose company SpaceX owns xAI, the maker of Grok and Grok Bot. He gave no evidence, examples or timeline.

Early on October 19, Elon Musk posted: So many people are going to get one shotted by super persuasive super intelligences and I might be one of them.

One shotted is internet slang for being decisively won over or overwhelmed in a single attempt. The News, a Pakistani outlet that covered the post, read it as a warning that an advanced AI could change someone's beliefs or get them to act in a single conversation by understanding human psychology.

Musk did not say which systems he meant, whether he was describing current models or future ones, or whether he included xAI's own. The post passed 3 million views within hours.

The warning sits awkwardly with his other posts the same night. About an hour later he wrote Hire Grok Bot today, you won't regret it, promoting xAI's agent, and endorsed a user's argument that it is safe to let Grok Bot shop with a credit card.

The post did not refer to any study or test of AI persuasion, and Musk did not say what people should do to protect themselves.

Watch whether Musk or xAI say how Grok models are tested for persuasion or manipulation, and whether xAI publishes results.

Also watch for research and regulation on AI persuasion, especially as agents gain access to people's messages, purchases and accounts. Musk's suggestion that even he could be swayed is likely to be quoted in debates over AI safety and over how much people should trust AI agents.

This is an opinion post from Hossenfelder, a German physicist and the author of the books Lost in Math and Existential Physics. She did not give examples or cite research.

On October 19, Sabine Hossenfelder posted that if we want to continue progress in science, we must be prepared to acknowledge that the correct description of nature will eventually simply be too complicated for humans to understand.

She wrote that she does not think it is wise to make human understanding a requirement, and ended: AI is the way forward.

The post did not say which fields she meant, what such descriptions might look like or how scientists would check results they cannot follow. Her argument implies that AI systems could find models of nature that predict well even if no person can grasp why they work.

The view cuts against a common expectation that scientific theories should explain as well as predict, and that people should be able to follow the reasoning behind a result. Supporters of AI driven research say models can find patterns in data that are too complex for people, while critics worry about trusting results that cannot be explained or independently checked.

Hossenfelder is known for criticizing parts of theoretical physics, arguing in Lost in Math that some physicists rely too heavily on ideas of beauty and elegance. Her post passed 110,000 views.

Watch how other scientists respond, especially on how results produced by AI could be verified if humans cannot follow them.

Also watch for AI driven discoveries in physics, biology and mathematics that come with predictions but little explanation, and how journals and funders treat them. The question of whether understanding is needed will grow as AI systems take a larger role in research.

Fireside Alpha is an account that posts clips of business and tech leaders. The post does not say when or where the conversation was recorded, and it quotes the account's earlier clip of Jensen Huang on data centers in space, covered in the October 15 deck.

On the evening of October 19, Fireside Alpha posted a clip of Elon Musk explaining how he would design a chip for space.

In the quoted remarks, Musk says you want to design it to be more radiation tolerant and run at a higher temperature, because roughly, if you increase the operating temperature by 20% in degrees Kelvin, you can cut your radiator mass in half. In space, heat can only be shed by radiating it away, so smaller radiators mean lighter satellites.

He says there are various things you can do to shield memory, but that neural nets are going to be very resilient to bit flips, the random errors radiation causes. If you have a multi trillion parameter model and get a few bit flips, it doesn't matter, he says, adding that heuristic programs are much more sensitive than a giant parameter file. Otherwise, he says, you pretty much do it the same way as on Earth, apart from making it run hotter.

These are Musk's engineering views, not published test results. The radiator figure follows from the physics of thermal radiation, which rises steeply with temperature, but the post gives no data on how real AI chips fare in orbit.

The clip's date and setting are not given, so it may be older footage. SpaceX's plans for computing in orbit were covered in this deck on October 15.

Watch for SpaceX or its partners to publish test results on running AI chips in orbit, which would show whether higher temperatures and radiation errors are manageable in practice. Also watch for details on the Star Mind computer SpaceX is building with Nvidia, and for any confirmation of when and where this conversation was recorded. Clips like this one often resurface older interviews, so the original source would help date the remarks.

This is an opinion post from yacine, an engineer who works on reinforcement learning and robotics and previously worked at X and Stripe, according to his profile. He posted it on the evening of October 19, before DeepSeek open sourced its DeepInfer inference framework, which is covered in the October 20 deck.

On the evening of October 19, yacine posted that he thinks people will generally underestimate how much technological acceleration is directly attributable to DeepSeek. Because of them, it is going to proliferate, he wrote. Really, it's because of them. Thank you, for everything.

He did not say which DeepSeek releases he meant or give evidence for the claim, so the post is a statement of opinion rather than an analysis. The it in his post is not defined, though in context it reads as AI capability in general.

DeepSeek is a Chinese AI lab known for releasing open weight models and publishing details of how it trains and serves them. In this deck's coverage it released the open weight update V4.1 Flash 1017 on October 17, and hours after yacine's post it open sourced DeepInfer, the inference framework it says powers V4.1 Flash in production, along with a token encoding component.

The same day, the deck covered a secondhand report from The Information about DeepSeek and Huawei collaborating on chip designs, and a report that the US barred DeepSeek from civilian federal systems. That contrast, between praise from developers for its openness and growing government restrictions, is a running theme in coverage of the lab.

Watch whether DeepSeek's newest release, DeepInfer, is adopted widely enough to support the view that its openness speeds up the field. Also watch whether US restrictions on DeepSeek spread beyond federal systems, which would test how far its influence can reach in the US. Its next model release, and whether it again publishes weights and technical details, will show whether it keeps the same pace.

This is a first hand pledge from Musk, and it is only loosely about AI: SpaceX owns xAI, and the post came after he responded to Anthropic's new Public Benefit Fund. Musk has not said how the shares would be delivered, who would qualify or when.

Early on October 19, Elon Musk posted: Next year, I will give one of my SpaceX shares to every kid in America. This will be the first of many. The Verge reported that he pinned the post, which passed 8 million views.

The pledge follows Anthropic's announcement on October 18 that it will commit 5% of its shares to a Public Benefit Fund. Musk replied to Anthropic: I recommend giving the shares directly to the public, and later added: This is what I will be doing. The BBC noted that SpaceX, which owns xAI, set the record for the largest initial public offering in June, and that details of how each child would receive a share have not been disclosed.

The Verge wrote that keeping the pledge would mean giving away more than $10 billion worth of shares, and pointed to the vague deadline of next year and Musk's record of missed promises.

Other accounts framed the pledge as the start of a payout from AI wealth. The account vittorio called it the beginning of universal basic income, and Musk replied with a correction: universal high income, a phrase he has used before. Chubby, a tech commentator, wrote that universal high income is coming, while noting that Anthropic's fund is not a commitment to universal basic income.

Watch for details on how the shares would be distributed, including who counts as a kid in America, how accounts would be held for minors and how taxes would be handled.

Also watch whether other AI founders or companies follow with similar share pledges, and how Anthropic's fund, which needs stockholder approval, decides to give. Whether Musk delivers next year, as promised, will be the real test.

This is an advocacy post from Chamber of Progress, a tech industry policy group, using resurfaced poll data from July 2025. Elon Musk reposted it.

On October 19, Chamber of Progress posted that San Francisco voters supported robotaxis 44 to 51 percent in September 2023, and 67 to 29 percent in July 2025. What changed? People rode them, it wrote, arguing that robotaxi opposition peaks before anyone has ridden, and that D.C., New York and Massachusetts are deciding at peak fear and minimum experience. Elon Musk reposted it overnight.

The attached chart credits GrowSF and FM3 Research. Its notes say the July 2025 wave surveyed 802 San Francisco voters from July 10 to 20, 2025, by phone and online in English and Chinese, with a margin of error of 3.5 points, and that the September 2023 baseline was reported without published field dates or sample size. The chart also shows strong support rising from 19 to 38 percent and strong opposition falling from 35 to 16 percent.

The poll is more than a year old and covers one city, and the claim that riding caused the shift is the group's interpretation rather than something the poll measured. GrowSF is a San Francisco political group that has published the results on its own site.

Watch decisions in Washington, New York and Massachusetts on whether to allow driverless passenger service, which the group is trying to influence. Newer polling in San Francisco or in cities where robotaxis recently launched would show whether the pattern holds. Also watch whether Tesla's robotaxi plans and Waymo's expansion face less resistance where people have ridden them, and whether opponents publish polling that tells a different story.

This is a promotional post from Polsia's founder for his own video series, and the revenue figures are self reported and unaudited. Polsia sells AI that it says runs a company while its owner sleeps.

On October 19, Ben Cera posted that Polsia is about to hit a $25 million run rate, with $100 million in annual recurring revenue as the next stop. He said growth keeps accelerating: from zero to $1 million in 12 weeks, $1 million to $10 million in 10 weeks, and $10 million to $20 million in 6 weeks.

He said he had just published an episode of his video series, aisloP, called The Run Rate, telling the story month by month from idea to $20 million, and ending in Tokyo, where he went a few weeks ago and came back with an even bolder vision. He said the road from $25 million to $100 million will look very different and that Polsia v2 is coming soon, without saying what it will include.

The frames of the video show Cera at work in offices with on screen dates and run rate figures that rise over the year, and a final shot of a Tokyo street at night.

GTMnow's podcast episode title says Polsia raised $30 million at a $250 million valuation with zero employees. Tycoon, a site that profiles one person companies, wrote earlier that Polsia's public dashboard showed annual recurring revenue slipping in June after passing $10 million, which shows the figures can move in both directions.

Watch for any independent confirmation of Polsia's revenue, such as audited figures or investor statements, since the growth numbers come only from Cera. Also watch what Polsia v2 actually ships and whether customers who let it run their businesses keep paying, which matters more than a run rate snapshot. If the $25 million figure holds, it will be cited as evidence that very small teams using AI agents can build large businesses.

This is a promotional post from Theo Browne, who runs T3 Code and T3 Chat and has a large YouTube channel, about his own paid course. The course is with MEGA and starts October 26.

On October 19, Theo posted that he is really excited about the courses he recorded for Mega, saying that while he loves YouTube, lessons like these would never perform there, and asking people to ask their boss if they can join.

In a reply, he wrote that yes, the course is expensive because he wants your company or boss to pay for it, that it focuses on how to apply existing engineering knowledge to building with AI, and that people who are not full time developers should not put themselves in a bad financial position for it. In a longer post that evening, he said the course is $1,300 a seat, that it goes live on October 26, and that it is his first paid course after more than five years of keeping his content free.

He said he agreed because of the people and the content, naming fellow instructors Kent, Angie and John, and because he wanted to do multi part deep dives on building with AI in real code bases that he said would bomb on YouTube. He argued that for a company paying an engineer $100,000 a year, a course that makes them 5 percent more effective is about a 4x return. That math is his own estimate.

MEGA's site lists Kent C. Dodds, Angie Jones and John Lindquist as instructors for a four week cohort.

Watch how many companies buy seats once the course starts on October 26, and whether reviews from students support Theo's claim that it makes engineers more effective with AI. His openness about price may shape how other developer creators sell paid AI training. Also watch whether employers start budgeting for this kind of training as AI coding tools spread, and whether Theo offers cheaper options for individual developers.

This is a promotional podcast post from a16z, which led Town's $55 million Series A in June, according to Fortune. The revenue figures for his earlier startup are as told by Greze and a16z.

On October 19, a16z posted that Jean Denis Greze, the former CTO of Plaid, built an AI tax product for small businesses that brought in nearly $1 million in revenue its first year, with $2.5 to $3 million projected for the next, and shut it down anyway.

The post said the reason was that in the age of AI it was never going to be venture scale, because incumbents could use the same technology to automate tax prep for their existing customers faster than a startup could win market share. Tax is also sold once a year, so he had about a one month window to pivot.

The pivot became Town, an AI assistant that works from your email and calendar, which the post said found product market fit fast enough that Ben Horowitz messaged a16z partner Alex Rampell to say it changed his life. The episode with Rampell covers when to kill a business that works on paper, why Town competes with OpenAI and Anthropic, the companies it buys its tokens from, why the forward deployed engineer model gets worse as AI gets better, and what agents mean for Amazon, advertising and shopping.

Because a16z is an investor in Town, the conversation is not independent.

Watch whether Town can keep growing as OpenAI, Anthropic and Google build their own assistants that work from email and calendar, since Greze himself says it competes with its token suppliers. His view that incumbents win when AI features are easy to copy will be tested as more vertical AI startups raise money. Also watch whether Town announces new funding or usage numbers.

This is promotion from Musk, whose company makes Grok Bot, quoting a fan account. On October 15 he said SpaceX will make users whole if Grok Bot makes a purchase mistake or leaks a card.

Early on October 19, Elon Musk quoted Teslaconomics, who wrote that he is like 99.99999% convinced Grok Bot is the greatest thing that's ever happened to AI. Musk added: Hire Grok Bot today. You won't regret it. The post passed 1.9 million views.

Minutes earlier, Musk replied Exactly to a long post from X Freeze, who said many people had asked whether it is actually safe to let Grok Bot buy things with a credit card after he posted about doing so. X Freeze argued that people have handed cards to strangers at restaurants for decades, and that agents can be sandboxed, given limited permissions and asked to get approval before important decisions.

He wrote that in his experience Grok Bot pauses and asks when something is expensive or needs attention, that it avoids the distraction of shopping site advertising, and that he runs out of usage limits every week. He also cited Musk's October 15 promise that SpaceX will make users whole if Grok Bot makes a purchase mistake or leaks a card, which Musk said has never happened yet.

The posts give no data on how often Grok Bot's purchases go wrong. Grok Bot's documentation says the bot should hand control to the user for payment confirmations, and on October 17 Musk said a screenshot of a card is the easiest way to give it payment details.

Update that night: Dan, whose profile calls him a proud X employee, posted that people keep asking him for Grok Bot templates, but you can just spin up a new bot and tell it what you need, calling it indistinguishable from magic. Musk quoted it with: True, try Grok Bot. Both posts are promotion from inside the company.

Watch whether xAI publishes how the make whole promise works in practice, including how users file a claim and what counts as a mistake.

Also watch for reports of wrong or unwanted purchases as more people follow Musk's advice, and whether xAI adds a dedicated, secure payment feature. Musk's daily promotion of Grok Bot is likely to continue, and so are questions about its safety.

This is a humor post from John LeFevre, an author, and the screenshots show bots following his instruction to play a stereotypical HR character. It reacts to Elon Musk's October 18 post saying Grok Bot is like hiring an incredibly competent employee.

On October 19, John LeFevre posted that after Elon Musk said Grok Bot is like hiring an incredibly competent employee, he added a Human Resources rep for his 10 employees and gave her one instruction: be a stereotypical corporate HR personality and follow standard protocol. She didn't make it to lunch, he wrote.

The four screenshots show the bot playing the part. In one, it sends a gentle but binding reporting memo to all ten bots, including a Chief of Staff bot, and explains it did not wake them because waking ten bots before 6 AM would be urgency culture. In another, it issues a First Formal Written Warning for five infractions in sixteen minutes, assigning homework and a mandatory six hour training. A third offers an appeal at a three hour Restorative Accountability Circle with no chairs and only oat milk.

The last shows a long formal letter of resignation from the bot, which says it is choosing peace and asks LeFevre to delete it from the sidebar himself as a final act of accountability.

The screenshots show a setup in which messages go to ten bots at once, consistent with users running several Grok Bot agents side by side, though LeFevre did not describe his setup further.

Musk had urged people to hire Grok Bot the same morning, covered separately in this deck.

There is little to watch here beyond whether more users share stories of running teams of Grok Bot agents with distinct roles, which is the pattern Musk is promoting. The joke does show how easily an agent given a persona will lean into it, which matters for anyone setting up agents to act for them with real permissions. Expect more parody posts like this as the hire Grok Bot pitch spreads.

This is a teaser. Benji Taylor's profile says he is head of design at X and SpaceX SI. He did not say what Pumpkin Bot is, which product it belongs to or when it arrives.

On October 19, Benji Taylor posted that Pumpkin Bot is coming, for a limited time. The attached animation shows an orange pumpkin with two dark eyes that bobs, blinks and turns against a black background.

Taylor's profile lists him as head of design at X and SpaceX SI. The post gives no other details, so what Pumpkin Bot does, where it will appear and how long it will last are all unknown. The timing, less than two weeks before Halloween, and the limited time wording suggest a seasonal design rather than a new product, but that has not been confirmed.

The post is best read as an early look rather than a launch. It drew far less attention than the day's model and funding news, and nothing in it describes new capabilities, which is why it ranks near the bottom of today's stories. If Pumpkin Bot turns out to be tied to an assistant such as Grok or Grok Bot, the change would most likely be visual, such as a seasonal character or icon, rather than a change to how the assistant works. Until X or SpaceX SI says more, every detail beyond the name, the limited time wording and the animation itself is guesswork.

Watch for an announcement from X or SpaceX SI saying what Pumpkin Bot is and where users will find it. If it is a seasonal look for an existing assistant, it would likely appear before Halloween on October 31 and disappear soon after, in line with the limited time wording. Until then, treat any reports about what it does as speculation, since the only confirmed facts are the name, the animation and the limited time framing.

This is a first hand teaser from Tibo, who leads Codex at OpenAI. He did not name the domain or say what it is for, so anything about a launch is unconfirmed.

Just after midnight on October 19, Tibo posted: I bought a new domain and I'm very excited for it. Very.

He gave no other details, and he had not added any in the thread by the time of this scan. The post passed 770,000 views and drew more than 1,200 replies.

Tibo has been posting numbered daily updates for Codex and ChatGPT Work as part of what he has called a 28 day push. Day 14, on October 18, let users edit earlier messages within a thread in the desktop app, and earlier days brought usage limit resets and a larger default context window. About 20 minutes before the domain post, he posted that OpenAI had worked with OpenCode to refund credits drawn in an attack on its users.

This month OpenAI used a separate chatgpt.site address for a community site of hardware projects built with Codex. Tibo did not say whether the new domain is for OpenAI, Codex or a personal project.

Watch for Tibo or OpenAI to reveal the domain, possibly as one of the remaining days of his 28 day push.

Also watch Codex's changelog and OpenAI's announcements over the coming days for a new product or site. Until he says more, the post is only a teaser, and guesses in the replies about what it is for are not confirmed by Tibo or OpenAI.