Andrew Curran is quoting a research post by William Zhou and Daniel Freeman of Anthropic's Frontier Red Team. The experiments ran in controlled settings, and the authors say the cyber exploit step was kept trivial, so results are an upper bound on real world success. The timeline is the authors' own projection.

On October 12, Andrew Curran posted that Anthropic's Frontier Red Team believes small AI models may now be months away from being capable of self replicating without access to a datacenter, quoting the blog's conclusion that the team is seeing signs of life for AI worms.

According to the post, the team ran Gemma4 26B across 15 off brand Android TV boxes costing about $55 each, at 0.39 tokens per second. In tests of setting up inference on unfamiliar hardware, a 9B model got inference running on 8 of 9 hardware types.

In a firewalled network of 8 cloud nodes running an Android emulator, with an unauthenticated ADB port as the entry point, GLM 5.2 was the only open weight model tested that completed end to end replication, reaching all 8 nodes in 4 of 5 trials and 7 in the other. The authors say models of 30B parameters and under can complete many steps but make mistakes chaining them, and that a trend line puts a small model at GLM 5.2's level by the end of 2026.

They estimate a replicator on a botnet the size of Kimwolf, over 2 million devices, would have at most about 40 H100s worth of inference and take days per replication, which they compare to a slime mold. They recommend ISP monitoring, quarantining infected customers and secure defaults from manufacturers.

Watch whether small open weight models reach the capability level the authors set by the end of 2026, as their trend line predicts.

Also watch whether ISPs and device makers act on the suggested defenses, whether regulators extend rules like the UK ban on default IoT passwords, and whether other labs publish similar replication evaluations for their own models. Independent replication of these results would also help.

Andreas Schilling is an editor at the German hardware site Hardwareluxx, and his post summarizes Intel's announcement at the Open Compute Project summit on October 12. Intel had outlined the Crescent Island architecture at Hot Chips earlier this year. The 480GB figure applies to partner built cards, while Intel's own card carries 160GB.

On October 12, Andreas Schilling posted the headline specs for Intel's GPU D Series, code named Crescent Island: 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth, 1.3 petaflops of FP8 and a 350 W TDP.

According to ServeTheHome, the first model is the GPU D270, built on Intel's Xe3P architecture with 32 Xe cores, 256 XMX engines, a clock speed of about 2.5GHz, 32MB of L2 cache and PCIe Gen5 x16. It is an air cooled PCIe card, and ServeTheHome noted that AMD's comparable MI350P runs at up to 600 watts in its top configuration.

TechPowerUp reported that Intel's branded card ships with 160GB, while ODM partners can fit up to 480GB, and that Intel says one card can run models of up to 300 billion parameters, with an eight card setup handling trillion parameter models. It added that the chip supports formats from FP4 to FP64 and includes four video encoders and four decoders.

Intel says customer sampling is underway with server makers, neoclouds and software vendors. Pricing, a launch date and independent performance results have not been disclosed, and ServeTheHome said buyers should look to 2027 for release.

Watch for pricing and a launch date, and for independent benchmarks of cost per token against Nvidia and AMD inference cards.

Also watch which server makers and neoclouds announce D270 systems, whether partners ship the 480GB configuration, and whether Intel's software stack is mature enough for customers to move inference workloads onto it. A 2027 launch would put it up against newer Nvidia and AMD parts.

The lead post is from ns123abc, a third party account, and is based on a screenshot of a terms of service page. The first hand confirmation came hours later from SpaceXSI's official account. Bloomberg reported in late September that the company was weighing a single Grok and X plan.

On October 12, ns123abc posted that SpaceXSI officially published the legal terms for a unified X Premium plus Grok plus Cursor subscription plan, calling X the everything app. The attached screenshot shows a page titled Bundled Subscription Terms of Service that refers to an X Subscription Consumer Terms of Service, a Grok Addendum and a Cursor Addendum.

At 7:08 PM ET the same day, SpaceXSI's official account posted: We've unified plans across Grok, Cursor, and X Premium. Introducing The X Subscription. The post included a video and did not list prices or tiers.

The Next Web, summarizing Bloomberg, reported on September 30 that a company document described four tiers, from a free plan with stricter Grok limits, through an $8 a month lite plan with a verified checkmark, up to a $100 a month Ultra tier that includes the Grok Bot agent. A leaked price card covered by CellCog listed different prices, so the final pricing remains unconfirmed until SpaceXSI publishes it.

The ns123abc post also mentions banking, but neither the terms screenshot nor SpaceXSI's post describes a banking feature.

Watch for official pricing and tier details for The X Subscription, and how they compare with the tiers Bloomberg described and the leaked price card.

Also watch how existing SuperGrok, X Premium and Cursor subscribers are migrated, whether Cursor stays available as a separate product, and whether the bundle pulls users away from rival AI subscriptions from OpenAI and Anthropic.

The underlying data comes from the Bureau of Economic Analysis GDP release, though the chart credits the Federal Reserve, whose FRED database republishes it. The 5.05% reading is for the second quarter of 2026, which BEA had already published, so this is not a new data release. The Kobeissi Letter is a markets commentary account.

On October 12, The Kobeissi Letter posted that US private fixed investment in information processing equipment and software is now up to a record 5.05% of US GDP, up 1.08 percentage points since the third quarter of 2023, and that the share peaked at 4.46% in the fourth quarter of 2000 before the 2001 recession.

FRED shows the series at a seasonally adjusted annual rate of $1,645.7 billion in the second quarter of 2026, matching the post's figure of $1.65 trillion. FRED shows $1,106.7 billion for the third quarter of 2023, a gain of about $539 billion, slightly more than the $531 billion in the post, possibly because of data revisions. Thetrading.tools, which computes the share from BEA data, also puts the second quarter reading at 5.05%, a record since 1959, up from 4.97%.

That site notes caveats. Data center buildings are counted as structures and are outside this series, though the servers and networking gear inside them are in it. Imported equipment counts here and is subtracted again in net exports, so a rising share overstates the boost to GDP growth. The category also includes medical instruments, which are unrelated to AI.

Watch the BEA's advance estimate for the third quarter of 2026 to see whether the share keeps climbing above 5.05%.

Also watch how much of the increase comes from imported servers and chips versus domestic spending, and whether tighter financing for AI infrastructure, which Broadcom's Hock Tan warned about on October 12, starts to slow equipment and software investment. Revisions to earlier quarters could also move the record.

Dexerto is relaying a Bloomberg report based on people familiar with the project and an internal message Bloomberg reviewed. The game is an unannounced shooter developed with Disney. Epic is the privately held maker of Fortnite.

On October 12, Dexerto posted that Epic Games has reportedly told devs to use more AI tools on its delayed Star Wars and Marvel shooter, and that one director told staff no more hand written code and said all game art should be generated with full AI workflows.

According to Quartz, summarizing Bloomberg, the game had been scheduled for release in March and now has no new date. Bloomberg reported that development lagged because Epic misjudged the headcount and timeline the project needed.

In an internal message Bloomberg reviewed, one of the project's directors told staff there would be no more hand written code and that the artwork would be generated with full AI workflows. Developers were told to use Anthropic's Claude, with humans managing and guiding the AI's output. Bloomberg said the news, shared on a video call last week, shocked a number of developers, who questioned how errors in AI generated work would be handled and doubted the plan would succeed.

Epic communications director Jake Jones told Bloomberg the company is embracing AI accelerated development led by Epic developers, whom he called the visionaries and creative drivers for everything Epic does.

Bloomberg described it as the first attempt by a major game company to build a title with AI as the primary foundation of its pipeline.

Watch whether Epic sets a new release date and how much of the shipped game's code and art is actually AI generated.

Also watch reactions from Disney, from game developer unions and from players, and whether other large studios follow Epic's approach or hold back after seeing how this project goes. Bloomberg's sources said staff doubted the plan, so any departures from the team would be telling.

The post does not cite a source; the quotes come from a Bloomberg report on Tan's remarks at an event in Singapore on October 12. Intel CEO Lip Bu Tan also spoke on the panel.

On October 12, unusual_whales posted that Broadcom CEO Hock Tan says some AI investors are betting like they're in a casino, quoting him: A lot of investors have no clue what they are investing in.

Bloomberg reported that Tan said investors in frontier large language models set very extreme expectations only to find out that it's almost like they are in a casino. He also questioned whether businesses are wasting money on AI and the longer term economics of money losing AI labs.

Tan said OpenAI reaching annualized revenue of $50 billion in September points to a viable business model but offers little guarantee of long term profitability, because it is still losing money on training. He said model developers need customers who earn returns from AI, not users wasting money on tokens, comparing it to a child spending $10 of arcade tokens in 10 minutes and learning nothing.

Tan warned that a shortage of capital could become the next constraint on growth. Bloomberg noted that Nvidia backed Firmus Grid withdrew a planned $5.5 billion IPO last week. Intel CEO Lip Bu Tan said hardware will stay in short supply for two to three years and that memory could be among the first segments to lose steam.

Watch whether other chip and cloud executives echo Tan's warning, and whether AI infrastructure IPOs and debt deals keep getting pulled or repriced after Firmus Grid.

Also watch Broadcom's own guidance on custom AI chip orders, since its business depends on the same AI lab spending Tan questioned, and whether OpenAI shows a clearer path to profit. Broadcom's next earnings call is the likely place for Tan to expand on these views.

Nvidia unveiled the RTX Spark superchip at GTC Taipei in May, and preorders for RTX Spark laptops opened at a Microsoft event in San Francisco on October 7. Nvidia says laptops ship October 16 and compact desktops follow in November. Only the Berlin event has published details so far.

On October 12, Nvidia posted that a new beginning for Windows PCs deserves a global celebration, inviting people to RTX Spark launch events in Beijing, Berlin, Istanbul, Seoul, San Francisco and Taipei to try demos, pick up swag and enter to win an ASUS ProArt P14 RTX Spark laptop signed by Jensen Huang. It said event details are coming soon.

TechPowerUp reported that the Berlin celebration takes place on October 16, the day the first laptops ship, at XPERION on Alexanderplatz. It said entry is free with registration, open to anyone 18 or older, and that Nvidia is lighting the Berlin TV Tower green. The event leads into GTC Berlin from October 20 to 22, with Huang's keynote on October 21.

According to Nvidia's blog, systems are coming from Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte. TechPowerUp reported that Microsoft's Surface Laptop Ultra starts at $2,599 and reaches $5,899 with 128 GB of memory, and that the ASUS ProArt P14 starts at $2,599.99 in the US.

Dates and venues for the other five cities had not been published at the time of the post.

Watch for dates and venues for the Beijing, Istanbul, Seoul, San Francisco and Taipei events, and for early reviews of RTX Spark laptops after they ship on October 16.

Also watch Huang's GTC Berlin keynote on October 21 for sales figures or new RTX Spark models, and how Intel, AMD and Qualcomm respond with pricing on their own AI laptops.

The Hormuz Letter is an aggregator account relaying a report by Tasnim, a news agency linked to the Islamic Revolutionary Guard Corps. The idea is not new: Tasnim and Fars floated cable permits and fees in May, and TIME then reported Iran had adopted no official policy or law on it. This follows the October 11 deck's story 23 on Mohsen Rezaei's threat against Gulf cables, data centers and AI facilities.

On October 12, The Hormuz Letter posted that Iran is moving to force Google, Meta, Microsoft and Amazon to obtain passage permits, undergo technical inspections and pay sovereignty fees for undersea internet cables running through its territorial waters in the Strait of Hormuz, per Tasnim.

Tamrah News, summarizing the Tasnim report, described it as a proposal that Tehran could require such licenses, inspections and fees, rather than a decision. It said the idea builds on Tasnim and Fars articles from May and that no later confirmation of an official policy was found. No company response to the latest report was found.

The post says at least two cable systems, FALCON and Gulf Bridge International, run through Iranian waters. Tamrah News, citing Marine Insight, reported the same, adding that most cables run closer to Oman, and TIME quoted experts saying operators routed most cables through Omani waters on purpose.

RFE/RL reported in May that experts see the proposals as more threat than workable plan, noting that the UN Convention on the Law of the Sea protects the right to lay and maintain cables and that Iran signed it but never ratified it. Tech Times reported that US sanctions bar American companies from paying Iranian state entities.

Watch whether Iran's government or parliament turns the Tasnim proposal into actual law or regulation, which did not happen after the May reports.

Also watch for responses from Google, Meta, Microsoft and Amazon, from the operators of FALCON and Gulf Bridge International, and from Oman and the UAE, and whether the threat is folded into talks over reopening the Strait.

Gregory Koberger founded ReadMe, a company that hosts API documentation for developers. Restless is a new product from the same team, and ReadMe's documentation calls it Restless by ReadMe. An API is the interface that lets software, including AI agents, call another company's service.

On October 12, Gregory Koberger posted: Your API is now your product, and it needs to move as fast as you do. Introducing Restless. Every call, support question, pull request, and error message leaves your API better for the next agent.

According to the Restless site, the product draws on real time API logs, answers given by human support staff, feedback from agents and existing documentation from ReadMe, Mintlify, Fern or other tools. It says it catches failed calls and tells the caller how to recover and retry, and that it adds a send_feedback tool to a company's MCP server so an agent can report what tripped it up instead of giving up.

ReadMe's documentation describes Restless as a modern API toolkit to make APIs agent ready. It says Restless runs in an API's request path, captures every API and MCP call, puts the fix right in the error response and gives teams one view of who is calling and what is failing. Setup starts with npx restless init, which installs the Restless SDK.

ReadMe says its customers do not need to migrate anything, can connect existing docs as a source, and get 4,000 extra credits a month. Restless did not publish pricing in the launch post.

Watch whether API companies adopt Restless alongside or instead of documentation tools like ReadMe, Mintlify and Fern, and whether agent feedback through MCP becomes a common pattern.

Also watch how teams handle putting a third party SDK in their request path, and whether Restless publishes data showing that its fixes actually raise the share of successful agent calls, along with the pricing it has not yet published.

Watcher.Guru's post is secondhand and names no source; the claim matches a Financial Times interview with Zuckerberg published the same day. Meta launched its Muse Wonder models on October 12. In September, Anthropic's Dario Amodei and OpenAI's Sam Altman called for coordinated efforts to pace AI development.

On October 12, Watcher.Guru posted: JUST IN: Mark Zuckerberg says $META has caught up with OpenAI and Anthropic in the AI race and rejects calls to slow down development. The post did not link a source.

The firsthand source appears to be the Financial Times, which published an interview the same day. According to the FT, Zuckerberg said Meta has caught up with rivals at the frontier of AI, committed to further vast spending to challenge Anthropic and OpenAI, and rejected calls to slow development. The full FT article is paywalled, so the deck does not quote it further.

The interview coincided with the launch of Muse Wonder and Muse Wonder Full. WIRED reported that Meta says the largest version beats Anthropic's Claude Fable 5.1 and ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index, while leading on some specialized benchmarks.

Zuckerberg's position on slowdowns is not new. In a September interview with NBC News, he said he does not think the industry needs some kind of industrywide coordination, and that each lab should take the time it needs internally when it sees issues. NBC reported that this cut against calls from Amodei and Altman for coordinated action.

Watch whether independent evaluations back Meta's claim of reaching the frontier, and whether Meta's promised safety report for Muse Wonder appears soon.

Also watch how Anthropic and OpenAI respond to Zuckerberg's rejection of a coordinated slowdown, and whether governments push the idea further after Amodei and Altman briefed the UN Security Council on pacing AI, and whether lawmakers in Washington take up any of their proposals.

OpenBMB is the open source group behind the MiniCPM family of small models, developed with Tsinghua University researchers and the company ModelBest. The new release has two models, a 35B mixture of experts model with about 3B parameters active per token and a 1B model. All performance figures are OpenBMB's own.

On October 12, OpenBMB posted that MiniCPM V 4.7 is its new open source multimodal LLM series, bringing strong multimodal capabilities to iOS and Android devices with extreme efficiency.

According to the post, the larger model uses a new technique called Lookahead Expert Scheduling, which predicts and preloads the expert weights the model will need next, to cut stalls during inference. OpenBMB says this gives a 5 times speedup on an iPad Pro M4 with 16GB of memory, with outputs identical byte for byte, and lets the model reach more than 17 tokens per second on that iPad and more than 13 tokens per second on an iPhone 17 Pro Max while using only 3GB of memory.

The post claims competitive results across 34 benchmarks covering multimodal understanding, STEM reasoning, OCR and long video comprehension. It says that even with 16 times visual token compression, the 35B model scores 90.4 on OCRBench, 5.2 points ahead of Qwen3.6 35B.

OpenBMB also introduces Self Replay Fine Tuning, which it describes as a low cost way to add new capabilities while keeping general performance, and says it has day zero support for Apple, Intel, AMD, Arm and Rockchip hardware. AIReport.NET reported the models are released under the Apache 2.0 license.

Watch for independent benchmarks and on device speed tests, since the 17 and 13 tokens per second figures come from OpenBMB.

Also watch whether app developers ship MiniCPM V 4.7 inside iOS and Android apps, whether Lookahead Expert Scheduling is adopted by other mixture of experts models for phones, and how the release compares with on device models from Apple, Google and Qwen.

Muse Wonder builds on Muse Spark, the first model from Meta Superintelligence Labs, released in April. Meta launched Muse, its personal AI agent app, on September 8. The benchmark comparisons are from Meta's own testing.

On October 12, Mark Zuckerberg posted: Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.

According to Meta's announcement, the series includes Muse Wonder and Muse Wonder Full, its most capable models to date. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. Meta said Muse Wonder will soon power more of its Muse agent, rolling out first to subscribers. The company described two features: Active Memory, which carries relevant details across conversations, and Active Focus, which keeps working toward a goal, recovers from errors and checks in before anything important.

The table Zuckerberg shared compares the models with Claude Fable 5.1, GPT 6 Astra and Claude Opus 5.5. Muse Wonder Full leads on several tests, including 59.1 on Meta's Personal Superintelligence Index and 79.0 on DeepSWE 1.1, but trails on others, such as ProgramBench, where Claude Opus 5.5 scores 70.5 against 59.0, and the Artificial Analysis Index, 56 against 58. The Verge noted it is unclear how the models compare with Google's Gemini 4 Argon.

Meta said a full safety report will be available soon.

Watch for independent benchmark results, especially from Artificial Analysis and other outside evaluators, and for Meta's promised safety report with additional evaluations.

Also watch how quickly Muse Wonder reaches Muse users beyond subscribers, how Meta prices the Model API against OpenAI and Anthropic, and whether the launch helps Muse hold onto users after its strong start, given Sensor Tower data showing ChatGPT overtaking it by Day 14.

Dexerto is relaying reports that began with two hardware leakers on X on October 10, which Tom's Hardware, TechPowerUp and others then wrote up. Nvidia has not confirmed the claim. The RTX 5090 uses the GB202 chip, the same silicon Nvidia puts in its far more expensive RTX PRO workstation cards.

On October 12, Dexerto posted that Nvidia has reportedly stopped RTX 5090 GPU production to prioritize more profitable AI chips used by data centers. The report is two days old and comes from leakers, not Nvidia.

Tom's Hardware reported that the leaker hongxing2020 cited a notice from an unnamed factory saying Nvidia will no longer manufacture any RTX 5090 models, including the D variants for China, and that an RTX 5080 with 24GB of memory will become the top GeForce card. It said a second leaker, MEGAsizeGPU, posted that all future GB202 supply is reserved for RTX PRO cards. TechPowerUp said Nvidia had not confirmed any of this.

Tom's Hardware argued the move would make economic sense, since the RTX PRO 5500 has the same 21,760 CUDA cores as the RTX 5090 but 84GB of memory, and a full RTX PRO 6000 can sell for about $15,000. It said a cheap 32GB RTX 5090 competes with Nvidia's own professional cards for AI work. The Threshold reported that as of October 11, Nvidia's US store still listed the RTX 5090 at $1,999, while Notebookcheck put street prices near $5,000.

Watch for any statement from Nvidia or its board partners confirming or denying the halt, and whether an RTX 5080 with 24GB of memory is announced.

Also watch RTX 5090 retail prices and stock over the coming weeks, since the reports say fresh supply would stop, and whether AI buyers shift to RTX PRO cards as a result. Any official word from Nvidia would settle whether this is a full halt or a quieter shift in chip allocation.

Z.ai is the international brand of Zhipu AI, the Chinese company behind the GLM model family. GLM Legal is available through an application for API access and is coming to AutoClaw, Z.ai's desktop agent platform. All benchmark results here are Z.ai's own and have not been independently verified.

On October 12, Z.ai posted that GLM Legal, built on GLM 5.3 Flash, matches frontier models on mainstream legal benchmarks and is designed for contract review, evidence analysis and compliance.

In its tech blog, Z.ai says GLM Legal scored a 19.2% all pass rate on Harvey LAB, a benchmark of real law firm workflows across 24 practice areas, an improvement of 12.5 percentage points over its base model. The chart in the post shows Kimi K3 at 12.9, Claude Fable 5 at 11.3 and GPT 6 Astra at 5.4 on the same test. Z.ai says the model matches top closed models on LegalBench and on PLawBench, a Chinese legal practice benchmark, despite a much smaller parameter footprint, though the blog does not give those scores.

The company says it trained the model with extra legal data in a mid training stage, then with reinforcement learning across tens of thousands of realistic legal environments, while improving the agent harness alongside the model. Its examples include contract review, analysis of 40 documents from a software dispute and privacy planning for a launch in Singapore, Malaysia and Indonesia.

Z.ai says GLM Legal is in beta for law firms, legal tech companies, in house teams and independent practitioners, and did not publish pricing or model weights.

Watch for independent results on Harvey LAB and LegalBench, and whether Z.ai publishes the LegalBench and PLawBench numbers behind its parity claim.

Also watch who joins the beta, whether the model reaches AutoClaw on schedule, and whether Z.ai releases weights as it has for other GLM models, which would let firms run it on their own servers. Independent tests by legal tech firms would be the clearest check on the parity claim.

Subquadratic is an AI research company that emerged from stealth this year with SubQ, which it called the first LLM built on a fully subquadratic architecture, where compute grows linearly with context length. Decision models return a choice, score or label with probabilities instead of open ended text. Jev is the decision model that SubQ compares itself against throughout its announcement.

On October 12, Alexander Whedon introduced SubQ Snap, calling it the first long context decision model for real enterprise work and the only decision model or LLM where compute per token stays constant as context grows.

He listed 384K tokens of multimodal context, 12 times larger than Jev; a score of 61.7 on Decision Index v0.2.1 against Jev's 57.9, outperforming on 13 of 16 enterprise benchmarks; decisions on 100K tokens in under one second; and pricing of $0.05 per million input tokens with free output. The first 10,000 sign ups get 100M free input tokens, and he warned latency will vary, giving speeds of 80ms at minimal context up to 2.5 seconds at 384K.

According to Subquadratic's press release, Snap runs on Subquadratic Sparse Attention and scores 96.3 on Decision RULER at 1M tokens and 74.0 on Decision LCR, against 48.0 for Jev with chunking. The release said Snap averages 67.6 to Jev's 61.6 across 16 Decision Index benchmarks, while the company's site lists 67.5.

The release quotes Whedon, identified as cofounder and CEO, saying today's decision models have context windows too small or slow for real enterprise work. All benchmark figures come from the company.

Watch for independent tests of Snap's long context accuracy and latency, especially whether constant compute per token holds up under real enterprise loads, and how fast Subquadratic adds capacity after warning it would hit limits.

Also watch how Jev's makers and other decision model teams, such as vLLM Semantic Router with Decision 3.0, respond on context length and price, and whether enterprise buyers adopt it for routing and screening work.

Getty Images supplies editorial and stock photos and video to media outlets and businesses. It abandoned a $3.7 billion merger with Shutterstock in late June after a UK regulatory demand, The Hollywood Reporter said. Zerohedge's line AI strikes again is its own commentary.

On October 12, zerohedge posted the headline Getty Images to go private via prepackaged Chapter 11, adding: AI strikes again.

According to Reuters, Getty said it secured support from major creditors and investors for a restructuring that will cut debt, raise fresh capital and take the company private. Affiliates of the Getty family and Koch Equity Development committed up to $533 million of new equity, holders of unsecured bonds agreed to retire $270 million of notes, and secured lenders agreed to lower interest costs and extend maturities. The investors will own 100 percent of the reorganized company, and Getty expects to finish by year end through an expedited prepackaged Chapter 11 for certain US entities.

The Hollywood Reporter said Getty expects no impact on operations, employees, partners or vendors. It reported that Getty disclosed in an investor presentation that its iStock eCommerce business has seen declining agency demand and disruption, which the publication said is presumably due to AI tools that make stock image creation easy. Getty projected its editorial unit, built on real news photos, will grow from $369 million to $400 million by 2029.

Getty's own statements cited debt and non operating legal liabilities rather than AI as the reason for the filing, so zerohedge's AI framing is an interpretation.

Watch for the Chapter 11 filing in the coming weeks and whether the court approves the plan by year end as Getty expects, and what the restructuring means for its legal liabilities.

Also watch Getty's strategy as a private company, including its copyright litigation and licensing deals with AI developers, and whether other stock image businesses face similar pressure from AI generated images.

This is a clip from the same Sourcery interview with Molly O'Shea covered in story 25 of this deck. The solopreneur is not named, so the figures cannot be checked independently. The clip adds one detail: the business automates newsletters for other people.

On October 12, Molly O'Shea posted a clip in which Perplexity CEO Aravind Srinivas says there is a one person company spending about $1M a year on Perplexity.

In the clip, Srinivas says there is one solopreneur, literally a one person company, spending upwards of a million dollars a year on Perplexity, and that this founder is making $3 million a year by automating newsletters for others.

O'Shea's post quotes her own earlier post with the full interview, which describes the spending as being on Perplexity Computer, the company's agent product. In that post she writes that Srinivas's thesis is that AI lets smaller teams with less capital build businesses, and that Perplexity is betting on local and hybrid inference on users' Macs instead of building data centers.

Reuters, citing The Information, reported in August that Perplexity's annualized revenue had passed $750 million, driven largely by Perplexity Computer. A customer spending over $1 million a year would be a small but notable share of that.

Perplexity has not named the customer or published how many customers spend at this level, so the example is an anecdote from the CEO rather than verified data.

Watch whether Perplexity names the customer or publishes figures on how many small businesses spend heavily on Perplexity Computer.

Also watch whether other AI agent companies report similar single founder customers, and whether automated newsletter businesses like this one hold up as email platforms and readers respond to more machine written content. Disclosure of the customer's identity would let others test the claim.

NVIDIA and Cisco announced USFP, short for Ultra Dense Small Form Factor Pluggable, at the Open Compute Project summit on October 12. It is a specification still in development, and no sampling or availability date has been announced. Pluggable optics are the swappable modules that convert electrical signals to light in switches and network cards.

On October 12, NVIDIA's newsroom posted that NVIDIA and Cisco are introducing pluggable optics to scale AI factories, saying USFP is the first module to support 400Gb/s per lane and delivers 4x the bandwidth density of today's optics.

According to the joint press release, the first planned USFP transceivers will pack 16 full duplex lanes into a footprint the size of an OSFP module, for 6.4Tb/s per module, with the ability to expand to 32 lanes and 12.8Tb/s. The modules are meant for both switches and SuperNICs. The companies say doubling both lane count and bandwidth per lane lets AI fabrics connect more 1.6Tb/s endpoints without adding network tiers.

The release says USFP uses a new two dimensional connector with a two step insertion design that engages the high speed contacts only after the module is seated, reducing wear and improving signal integrity. Cisco's blog says the format supports both 200G and 400G SerDes systems and offers about 1.3 times better power efficiency.

NAND Research noted that the 1.3 times efficiency figure has no independent measurement yet, that TE Connectivity and Amphenol are involved as connector partners, and that the specification has no availability date.

Watch for a published USFP specification, first module samples and an availability date, none of which have been announced.

Also watch which optics makers and switch vendors beyond NVIDIA and Cisco commit to the format, and whether it competes with or complements co packaged optics, which NVIDIA and others are also pushing for future AI networks. Power and cooling limits in large clusters will shape which approach wins.

The Open Compute Project is an industry group, started by Facebook, that publishes open hardware designs for data centers. MGX is Nvidia's modular reference architecture for building servers and racks around its chips. Vera Rubin NVL72 is Nvidia's next rack scale system, pairing its Vera CPUs with Rubin GPUs.

On October 12, NVIDIA AI Infrastructure posted that at OCP Summit 2026 Nvidia contributed the Vera Rubin NVL72 rack and compute tray designs to the Open Compute Project through the open third generation MGX architecture, and announced advances across open source software, AI networking, pluggable optics and facility design standards.

According to Nvidia's blog, the event took place at the San Jose McEnery Convention Center. Nvidia said the updated MGX 1.2 specification adds rack scale blind mate liquid cooling, interoperable rack scale interconnects and a liquid cooled 5,000A busbar for higher power systems, and supports rack deployments up to 250kW. It also contributed a design for a Vera CPU rack meant for tool calling, orchestration and sandboxed execution by AI agents.

The blog listed partner announcements as well: open AI factory design and grid resilience standards published with hyperscalers, new open source components for its DSX AI factory software, networking work with Meta on Spectrum X Ethernet, an 800V DC factory design from Visionbay, a new pluggable optics standard with Cisco called USFP, and Falcon Transport support with Google on future ConnectX 10 SuperNICs and BlueField DPUs.

Ian Buck, Nvidia's vice president of hyperscale and high performance computing, spoke about open infrastructure at the event, Nvidia said.

Watch which server makers and cloud providers ship Vera Rubin systems built on the contributed MGX designs, and whether the open specs reduce the time it takes to bring new AI data centers online.

Also watch how rival chipmakers such as AMD respond at OCP, since open rack standards built around Nvidia's architecture could make it harder for alternative accelerators to fit into the same data center designs.

The post is a one line alert based on Downdetector, which counts user complaints rather than confirmed failures. The post gives no detail on which services were affected, and OpenAI's own explanation is not linked in it.

On October 12, DeItaone posted: USER REPORTS SHOW PROBLEMS WITH OPENAI: DOWNDETECTOR. The post went up at 10:54 AM ET.

StatusGator, which tracks official status pages, lists a major disruption to ChatGPT and Codex that OpenAI acknowledged on October 9 and that ended at 4:06 PM UTC on October 12, about an hour after the post. StatusGator's summary says ChatGPT conversations were not responding, responses were slow or streaming failed, and Codex features were unavailable in many cases. A related entry describes unavailable content in the ChatGPT Android app, trouble managing plugins in workspaces and delayed cost data in the Compliance API.

A web search summary of an OpenAI community thread said OpenAI attributed the disruption to a change in web traffic routing configuration that overloaded a smaller set of capacity, but that explanation could not be confirmed directly on OpenAI's status page here.

StatusGator also lists smaller OpenAI incidents on October 6, 7 and 8, including problems with Codex, the desktop app and Workspace Agents.

Watch for a postmortem from OpenAI on the October 9 to 12 disruption, including the root cause and how many users were affected.

Also watch whether the string of incidents over the past week continues, and whether business customers relying on Codex and ChatGPT Work push for stronger uptime commitments or spread their work across other AI providers. OpenAI's status page is the best place to confirm whether the incident is fully closed.

Michael Smith, 54, of Cornelius, North Carolina, pleaded guilty in March to one count of conspiracy to commit wire fraud. Streaming services pay royalties per stream from shared pools, so fake streams take money from genuine artists. Polymarket posted the news on October 12, six days after the sentencing.

On October 12, Polymarket posted that a man was sentenced to 18 months in prison after generating hundreds of thousands of AI songs and using bots to stream them billions of times, outstreaming Taylor Swift and earning $8 million. The sentence was handed down on October 6.

According to the US Attorney's Office for the Southern District of New York, Judge John G. Koeltl sentenced Michael Smith to 18 months in prison and two years of supervised release, and ordered $8,091,843.64 in forfeiture. Prosecutors said that from 2017 to 2024 Smith used fake email accounts and fraudulently obtained debit cards to create thousands of bot accounts, at times as many as 10,000 at once, on platforms including Spotify, Apple Music, Amazon Music and YouTube Music.

The office said Smith's AI generated songs were streamed billions of times. In April 2023, it said, Taylor Swift's entire catalogue received 9.3 million family plan streams on YouTube Music, while Smith's bots generated 80.9 million family plan streams of his music.

Billboard reported that Smith's lawyers asked for no prison time, arguing no victim suffered real harm, while federal guidelines called for 46 to 57 months and probation officials recommended 24. Billboard called it the first criminal case over streaming fraud.

Watch whether prosecutors bring more streaming fraud cases now that a sentence is on the books, and whether platforms such as Spotify and Deezer tighten detection of AI generated tracks and bot listening.

Also watch how labels and artist groups use the case to push for rules on labeling AI music and on how royalty pools are protected from automated streams.

Molly O'Shea hosts the Sourcery interview show. The financial figures in her post are reported numbers rather than Perplexity disclosures. Perplexity's $20 billion valuation dates from a round in September 2025, and in August 2026 The Information reported talks with Nvidia at more than $30 billion.

On October 12, Molly O'Shea posted an interview with Perplexity founder and CEO Aravind Srinivas on why AI native SMBs could drive the next American millionaire economic boom.

Her post describes Perplexity as valued at about $20B with about $2B in reported funding, $750M in annualized revenue as of August and 5x year over year revenue growth as it expands from search into autonomous agents. Reuters, citing The Information, reported in August that Perplexity's annualized revenue had passed $750 million, and that Nvidia was discussing an investment at a valuation above $30 billion. TechCrunch reported the $20 billion valuation in September 2025.

According to the post, Srinivas shared the example of a one person company generating $3M in annual revenue while spending over $1M a year on Perplexity Computer. The company was not named, so the example cannot be checked. His thesis, per the post, is that AI lets smaller teams with less capital build businesses.

O'Shea writes that Srinivas sees a shift toward local and hybrid inference, with agents running across personal devices and the cloud, and that Perplexity is betting on users' Macs instead of building data centers. The episode also covers Nvidia, Apple Silicon and what the post calls the Hugging Face incident.

Watch whether Perplexity closes the reported round with Nvidia, and at what valuation, and whether it discloses revenue figures directly.

Also watch how Perplexity Computer is priced and adopted by small businesses, and whether its local inference approach on Macs holds up against rivals that run agents mostly in their own data centers. Naming the solo company behind the $3M example would let others check Srinivas's claim.

The post relays preliminary estimates from Counterpoint Research, a market research firm, which Reuters also reported on October 12. The figures are preliminary and could be revised.

On October 12, DeItaone posted that global smartphone shipments fell 20% in Q3, hitting their lowest third quarter level since 2012, according to Counterpoint Research, and that AI data centers are absorbing memory chip supplies.

Reuters, as republished by BusinessDay, reported that shipments fell 20% year on year because soaring memory chip costs and supply shortages pressured manufacturers and pushed up prices. Counterpoint senior analyst Shilpi Jain said consumers are holding on to their devices longer than estimated, increasingly considering refurbished devices and prioritising higher value devices.

According to Mexico Business News, which summarized the report, Samsung kept the top spot with 23% market share while Apple reached 22%, its first third quarter above 20%. Ground News summaries said retail prices rose about 20% as manufacturers favored profitability over volume, and that Samsung's own shipments fell 3%.

The post's figure of a Xiaomi decline of over 30% did not appear in the coverage reviewed here, so it is unconfirmed.

Watch for final third quarter figures from Counterpoint, IDC and Canalys to see whether the 20% drop holds, and for Xiaomi's own numbers.

Also watch memory contract prices for DRAM and NAND, holiday phone pricing, and whether phone makers keep cutting volume to protect margins while AI data centers absorb supply into next year. A rebound in the fourth quarter would suggest the drop was a supply squeeze rather than weaker demand.

The Financial Times published the report on September 22, and unusual_whales resurfaced it on October 12. Britain's AI Security Institute tests frontier models before release under voluntary agreements with developers. Several senior safety researchers at leading labs have quit in recent years citing concerns about safety.

On October 12, unusual_whales posted that AI staff are complaining of mental toll over fears of threats to society, per FT. The Financial Times story it refers to ran on September 22.

The FT reported that multiple staff at the UK's AI Security Institute have been signed off work with stress and are undergoing counselling, according to people with knowledge of the matter. Four people told the paper that tight schedules for testing unreleased models, along with alarm over the rapid rollout and growing capabilities of AI, have contributed to low morale and burnout at the institute.

The FT said staff at Anthropic, OpenAI and Google DeepMind have also publicly complained of similar issues, and that more than a dozen senior researchers across the three labs have quit over the past two years citing concerns about how their leaders prioritize safety. It reported that a DeepMind employee described making AI safer as feeling like an impossible task.

According to the Daily Mail's summary of the FT report, a former institute employee said staff were particularly alarmed by a recent jump in capabilities related to cyberwarfare and biochemistry, and a DeepMind employee said there had been various quittings and a lot of mental health leave.

Watch whether the AI Security Institute or the UK government responds with changes to staffing or testing timelines, especially now that the institute is regaining early access to US models.

Also watch whether labs disclose more about staff wellbeing and attrition on safety teams, and whether more researchers resign publicly, which has repeatedly shaped the debate over how fast frontier AI should be released.

DHH, David Heinemeier Hansson, created the Ruby on Rails framework and cofounded 37signals, the company behind Basecamp. He was quoting a January post by Wes Roth with a clip of Dario Amodei at the World Economic Forum in Davos. Amodei is the CEO of Anthropic, which makes the Claude Code coding agent.

On October 12, DHH quoted a January post about Dario Amodei's Davos prediction and wrote: Dario turned out to be right. Maybe even a little cautious. He added that the replies at the time were all so very convinced that they knew more and better, and what an idiot he was.

According to Hindustan Times, Amodei said in a World Economic Forum session with Google DeepMind CEO Demis Hassabis that he had engineers at Anthropic who no longer write code and instead let the model write it and edit it. He said, I think, I don't know, we might be six to twelve months away from when the model is doing most, maybe all of what SWEs do end to end, and stressed that there was a lot of uncertainty.

That was in January, so his window runs to roughly now. DHH did not offer data, and whether models today do most of what software engineers do end to end is still disputed. Many teams report AI writing a large share of their code, while engineers still set requirements, review output and make design decisions.

The post stands out because DHH has often been skeptical of industry claims, and he is now crediting a forecast that many developers mocked.

Watch for hard evidence on whether AI agents now handle most software engineering end to end, such as company disclosures of AI written code, hiring data for engineers and results on realistic benchmarks.

Also watch whether Amodei or Anthropic claim the prediction has come true, and how Hassabis, who expected a slower timeline at Davos, and other skeptics respond as the year closes.

QwenGyre is described in an arXiv paper from Alibaba Group with researchers from USTC and Tsinghua University. It targets online reinforcement learning, where a model improves by attempting tasks and being scored. All results are the authors' own and have not been independently reproduced.

On October 12, Alibaba_Qwen posted that QwenGyre brings elastic reinforcement learning to extreme long horizon agent training.

According to the paper, a single agent run in this regime can span hours, hundreds of interactions and nearly 1 million tokens. The authors say two problems follow. Run times vary widely, so GPUs sit idle waiting for the slowest runs, and agent tools that compact context, call sub agents and retry create branching histories with large redundancy.

QwenGyre's elastic scheduler moves groups of GPUs between generating rollouts and training as demand shifts, without interrupting live runs. Its trajectory processor turns branching runs into trees with shared prefixes, scores partial progress after a timeout and caps the number of training samples per run.

The paper reports speedups of up to 1.85 times over a colocated baseline and 1.78 times over an asynchronous one. On the flagship Qwen 3.8 2.4T model with about 700,000 tokens per rollout, it says NL2RepoBench rose from 52.5% to 58.5% in 48 steps, and 48 steps took 75.42 hours versus 134.55 hours for the asynchronous setup. The experiments ran agents through Claude Code as the harness.

Watch whether Alibaba releases QwenGyre code, since the post links only the paper, and whether other labs adopt elastic GPU allocation for long agent training.

Also watch whether the gains on Qwen 3.8 2.4T show up in future Qwen releases, particularly on coding benchmarks such as NL2RepoBench, and how other research groups' results compare at smaller GPU budgets, which the paper says it did not test.

Maia is Microsoft's family of custom AI accelerators, and Maia 300 is its third generation. The Verge reported that Maia 200 was announced only in January and already runs production workloads including Microsoft Copilot, GitHub Copilot and Microsoft Foundry. Microsoft compares Maia 300 against leading GPU based systems without naming a specific product. Microsoft also announced the chip first hand in its own post on X, saying Maia 300 delivers 40% more inference tokens per megawatt than leading GPU based systems and up to 7 times the rack scale token throughput of Maia 200.

On October 12, Satya Nadella called the announcement Microsoft's biggest bet yet on its own silicon. He said that starting next calendar year Microsoft will scale its Maia family to multi gigawatt deployments, making it the primary inference system for its own products and services, and introduced Maia 300, which he said delivers 40% more tokens per watt than leading GPU based systems.

In the official Microsoft Blog post, cloud chief Scott Guthrie framed the figure as 40 percent more tokens per megawatt, and said a full 96 accelerator Maia 300 configuration delivers up to 7 times the rack scale token throughput of Maia 200. According to the post, each rack holds 96 accelerators within a 167 kilowatt power envelope, with 27.6 TB of HBM and 1.7 PB/s of aggregate memory bandwidth. Maia 300 will begin production deployments in 2027.

Microsoft's Azure architecture blog said Maia 300 offers more than 2.5 times the performance per watt of Maia 200 and twice the accelerators per rack, and is designed for long context reasoning, latency sensitive conversation and multi step agentic tasks.

Nadella added that Maia 200 is already deployed across the US, Europe and Australia, delivering up to 1.7 times more tokens per watt in Teams and Copilot. The Verge said the gains should let Microsoft rely mainly on its own chips instead of ones from Nvidia and AMD.

Watch for independent measurements of Maia 300 once it reaches production in 2027, since all efficiency figures so far are Microsoft's own projections, and for how much of Microsoft's GPU spending shifts as a result.

Also watch for more details at Microsoft Ignite next month, whether Maia becomes available to outside Azure customers at scale, and how Nvidia and AMD respond as another major buyer moves inference onto in house silicon.

LinkedIn's annual Jobs on the Rise list ranks the fastest growing roles in the US over the past three years, based on its own member data. The 2026 list was published in January, and unusual_whales shared the finding on October 12. AI engineers ranked first.

On October 12, unusual_whales posted that AI consultant and strategist is the second fastest growing role in the US, per LinkedIn. The ranking comes from LinkedIn's Jobs on the Rise 2026 list, published in January, not from a new report.

According to LinkedIn, AI consultants and strategists help organizations plan and implement AI technologies to improve operations and reach their goals. Their most common skills are large language models, MLOps and computer vision, and the most jobs are in San Francisco, New York City and Boston. LinkedIn said the role is 18 percent female and 82 percent male, with a median of 8.2 years of prior experience, and that people most often move into it from founder, software engineer and product manager roles. It said 30.3 percent of the jobs are remote and 32.8 percent hybrid.

AI engineers topped the list, followed by AI consultants and strategists, new home sales specialists, data annotators and AI and ML researchers, according to Business Insider. LinkedIn senior director Laura Lorenzetti told Business Insider that the ranking shows a whole adjacent system of how you implement AI, including culture change and getting people to adopt it.

LinkedIn also highlighted a rise in founders and independent consultants as a shift toward self employment.

Watch for LinkedIn's 2027 Jobs on the Rise list in January, which will show whether AI consultant roles keep growing or level off as companies build their own AI teams.

Also watch whether demand for AI adoption work holds up if AI agents take over more of that integration themselves, and whether entry level paths into these roles open up, given the median of eight years of prior experience.

Small World in Motion is Nikon's annual competition for video taken through microscopes. Nikon announced the decision in a statement dated October 9, and the BBC, The Verge and Nature reported it that day, so CNN's October 12 post is a later write up. CNN calls it a photography competition, but the disqualified entry was a video.

On October 12, CNN posted that the winner of a prestigious scientific photography competition has been disqualified after organizers concluded that his use of generative AI broke the rules.

Nikon said that after reevaluating the video and supporting materials and consulting its judging panel, it determined the winning video did not comply with the competition rules on generative AI. It said the decision should not be read as a judgment of the entrant's professional reputation, scientific contributions or intent.

According to Nature, the video by Ning Xu, an optical engineer at the National University of Singapore, showed cilia moving in the lungs and won first place on September 15. Within days, scientists commented on LinkedIn that some features looked biologically implausible. Nature reported that Xu wrote that his team used an AI model in post processing to enhance the regions below the cilia without making anatomical claims about them. The BBC described Xu as being from Tsinghua University and said he believed he had acted within the rules.

The Verge reported that a video by Nguyen Nam Nhat now holds first place, and that Nikon plans to revisit its rules and evaluation procedures.

Watch for Nikon's revised rules for next year's Small World contests, especially how they define acceptable AI assisted processing versus generated content.

Also watch whether journals and other science imaging competitions adopt similar disclosure requirements, and whether Xu or his institution publish the raw data behind the video to settle questions about what it shows. The outcome will shape how researchers present AI enhanced imaging in future entries.

The New York Times and other news publishers are suing OpenAI and Microsoft for copyright infringement over AI training, in a consolidated case before Judge Sidney Stein. The quote comes from a brief that was unsealed on September 17, and unusual_whales resurfaced the Washington Post's report on October 12. Microsoft has argued in court that training on such material is fair use.

On October 12, unusual_whales posted that Microsoft's director of applied science has called the training of AI models the largest theft of labor in human history, citing the Washington Post. The Post first reported the quote on September 17, when a legal brief filed by the New York Times was unsealed.

According to TechCrunch and Ars Technica, the executive is Brent Hecht, Microsoft's Director of Applied Science. In a January 2023 internal memo he called AI training through mass scraping an astonishing theft of unprecedented proportions and the largest theft of labor in human history. Ars Technica reported that he also wrote the plan to widely scrape news made a complete mockery of the idea of fair use.

TechCrunch reported that the filings say OpenAI's mid training datasets alone contain more than 91,692 copies of works from the Times, the Daily News and the Center for Investigative Reporting. According to It Does What Now and other outlets, a later Hecht memo warned that Copilot risked becoming a doom loop for the publishers whose content trained it.

Microsoft told reporters the memos reflect Hecht's personal opinion, not the company's legal position.

The Next Web reported that Microsoft's own data showed Copilot cut news click throughs by up to 94 percent.

Watch for Judge Stein's ruling on the summary judgment motions in the consolidated case, which could decide whether training on news content counts as fair use.

Also watch whether Microsoft or OpenAI respond to the internal documents in their own filings, and whether the revelations push more AI companies toward paid licensing deals with publishers, which Satya Nadella reportedly endorsed for paywalled content in his deposition.

Phoronix is reporting on a Git tag message Torvalds wrote, so the quotes are his own words. It follows an earlier Phoronix report that Torvalds used Claude Code to draft a separate kernel improvement. The patches had not yet been merged to the main branch when Phoronix wrote.

On October 12, Phoronix posted: Linus Torvalds "Did What I Preach", Used AI To Make Code Decipherable & Uncovered More Bugs, noting it was more AI use on the Linux kernel by Torvalds that weekend.

According to Phoronix, an assoc array fixes tag appeared in Torvalds' source tree on Sunday alongside the delayed Linux 7.3 rc7 release. In the tag message, Torvalds wrote that Nils Ollrogge reported a bug in the assoc_array library and had a patch to fix it, but that the code was too obscure for him to follow, so he did what he preaches and used modern LLM tools to help make sense of it all.

Torvalds wrote that the result is an LLM assisted branch that makes the code more readable by splitting up some larger, hard to read functions. He said the code was then run through a test harness in user space, which confirmed the reported bug and found three other bugs too, all going back to the first commit adding the library in 2013. The assoc_array code is the kernel's associative array implementation.

Phoronix said the patches will presumably go into the Linux 7.4 merge window, and that Torvalds has been increasingly vocal about the benefits of AI as a tool for kernel development.

Watch whether the assoc array fixes land in the Linux 7.4 merge window, and whether other maintainers follow Torvalds in using LLMs to refactor and test older kernel code.

Also watch the kernel community's proposal, reported by Phoronix, to drop the requirement for an assisted by tag on AI generated contributions, and how it balances disclosure with wider AI use.

BBCWorld is linking its own interview: Smith spoke to Will Bain on the BBC's Big Boss Interview. Polymarket posted the quote earlier, on October 11.

On October 12, BBCWorld posted: AI is less regulated than milk, says Microsoft president.

According to the BBC, Smith said: Right now, your washing machine is subject to more safety rules than AI. Every carton of milk that you see in the grocery store is subject to safety rules and inspection. He called AI both a tool and a weapon that urgently needs regulating, and said AI is moving forward faster than the speed of safety.

Smith set out three measures. Firms need emergency brakes, often called kill switches, to pause or slow AI that gets out of control. There should be a legal requirement to write incident reports and tell victims when a rogue AI agent breaches their systems. The biggest question, he said, is when government evaluators should have to sign off on frontier models before release.

The BBC reported that Smith said the UK is best placed to lead, calling its AI Security Institute the strongest in the world at testing models. He said industry self regulation is a good and necessary step but no longer sufficient, because the companies that are not as responsible are the ones to worry about. The BBC noted his views differ sharply from President Trump's push to go full steam ahead on AI.

Watch whether the UK moves toward mandatory testing or sign off for frontier models, and whether other AI companies back or reject Smith's proposals.

Also watch for any US response, given the administration's stance against heavy AI rules, and whether incident reporting or kill switch requirements appear in upcoming UK or EU measures. Smith's comments also put pressure on rival AI companies to say where they stand on government testing.

The forecast comes from analyst Kim Sun woo at Meritz Securities, according to Korean press reports. The post's claim that he has the best track record for forecasting SK hynix earnings is the poster's own description and is not verified here. SK hynix has not yet reported third quarter results.

On October 12, jukan05 posted that the Korean analyst with the best track record of forecasting SK hynix earnings released a report predicting a substantial beat versus consensus, with operating profit of KRW 80.1 trillion against a consensus of KRW 74 trillion.

Chosun Biz reported that Meritz Securities analyst Kim Sun woo expects third quarter operating profit of 80.1 trillion won, up 32% from the previous quarter. Meritz kept a Buy rating and raised its target price to 3.2 million won, against a previous close of 1.681 million won. Chosun Biz said recent estimates had fallen to 70 trillion won, partly because of the strong won.

Kim expects DRAM shipments up 13% with prices up 22%, and NAND shipments up 4% with prices up 15%. He forecast 63 trillion won of operating profit from DRAM and 17 trillion won from NAND, an operating margin of about 80%. He said HBM4 supply is starting at full scale, and that long term contract talks with key mobile customers are shifting toward terms favorable to suppliers.

Meritz raised its full year forecasts to 264 trillion won for this year and 454 trillion won for next year, and Kim said a memory shortage in 2027 is clear.

Watch SK hynix's actual third quarter results later this month against the 80.1 trillion won forecast and the 70 to 74 trillion won range of other estimates.

Also watch HBM4 shipment volumes, the outcome of long term supply talks with phone makers, and whether a strong won keeps weighing on reported profit, since all three drive how much of the AI memory boom shows up in earnings.

In January 2024, IMF staff analysis found that almost 40 percent of global employment is exposed to AI, meaning AI could affect tasks in those jobs. The IMF said roughly half of exposed jobs in advanced economies may benefit from AI, while the other half could see lower demand, lower wages or reduced hiring. The IMF and World Bank annual meetings opened in Bangkok this week.

Early on October 12, Watcher.Guru posted: JUST IN: IMF predicts AI will take over 40% of all jobs. The post cited no report, and the IMF has not said that.

The 40 percent figure comes from IMF staff research published in January 2024, which found that almost 40 percent of global employment is exposed to AI. In advanced economies about 60 percent of jobs are exposed, and the IMF said roughly half of those may benefit through higher productivity, while for the other half AI could take over key tasks, lowering labor demand, wages and hiring. Only in the most extreme cases, it said, might some jobs disappear. Exposure was put at 40 percent in emerging markets and 26 percent in low income countries.

The post came as IMF Managing Director Kristalina Georgieva repeated that research at the annual meetings in Bangkok. According to Reuters, she said on October 11 that some 40 percent of jobs will be either augmented, downgraded or eliminated by AI, and as high as 60 percent in advanced economies.

CNBC reported that on October 12 she described AI as hitting labor markets like a slow moving tsunami, with young graduates entering routine jobs most exposed.

Watch the IMF's updated World Economic Outlook, which Reuters said is due on Tuesday, for any new estimates of AI's effect on jobs and growth.

Also watch whether Watcher.Guru corrects or clarifies its post, and whether finance ministers at the Bangkok meetings back the retraining and safety net measures Georgieva and World Bank officials have urged for workers hit by AI.

Britain's AI Security Institute was set up after the first AI safety summit in 2023 and tests advanced models for capabilities and risk of misuse. According to the Telegraph, the White House asked Anthropic, OpenAI and other companies last month to stop sharing models with the institute. Reuters reported that the request was pending a US review.

Reuters reported on October 11, citing the Telegraph, that American AI developers can give the British government early access to models for safety testing without running afoul of the Trump administration. The administration reversed a request that developers such as OpenAI and Anthropic not share new models with Britain's AI Security Institute before release, the newspaper said.

The Telegraph reported that the request came from the White House Office of the National Cyber Director, and that after it Anthropic restricted UK officials from reviewing its new Claude Mythos 5.1 model. The paper said the decision was reversed after lobbying by AI minister Kanishka Narayan, and that the institute will get pre release access to models from Anthropic, Google DeepMind and OpenAI.

Reuters said the change signals a softer stance on international AI safety efforts, noting that Trump's new AI task force said on Friday that developers must report security incidents with their models. Spokespeople for Trump, OpenAI and Anthropic did not immediately respond to Reuters.

On Monday, according to Reuters, Narayan confirmed that institute scientists are now evaluating the latest generation of models, including Google's Gemini 4 Argon, under voluntary agreements with developers.

Watch whether Anthropic and OpenAI confirm that pre release access has resumed, and whether the institute gets to test Claude Mythos 5.1 and the next OpenAI releases before launch.

Also watch for the details of the US review that prompted the original request, and for how the new AI task force's incident reporting rules fit alongside voluntary testing deals with allies such as Britain.

Honor is a Chinese smartphone maker. According to Honor, it launched the Robot Phone in China on August 12, with sales starting August 18, after first showing it at MWC 2026 in March. The video in Nawfal's post comes from another user's X post.

On October 12, Mario Nawfal posted that China made a smartphone that can literally move its camera like a robot, saying Honor's Robot Phone has a 200 MP camera on a robotic arm that pops out of the back, tracks people, turns its head, nods and reacts to you through AI.

Honor says the device integrates a motorized gimbal into the phone. Its 200MP main camera, with a 1/1.28 inch sensor, sits on the gimbal and supports ARRI LogC3 video, alongside a 200MP periscope telephoto with 2.7x optical zoom and a 50MP ultrawide. Honor priced it at 9,999 yuan for 12GB and 512GB and 12,999 yuan for 16GB and 1TB.

TechRepublic reported that the camera works with Honor's Yoyo AI assistant, letting the phone observe its surroundings, recognize facial expressions and perform gestures such as nodding, peeking and moving to music. It said the starting price is about $1,480.

Engadget's hands on found the stabilization impressive but called the phone a proof of concept and an expensive early adopter experiment. It reported that the phone will not officially launch outside China, though an Honor spokesperson said the company is looking into future devices that could.

Watch whether Honor brings the Robot Phone or a successor outside China, which its spokesperson told Engadget the company is exploring, and whether rivals add moving camera hardware to their own phones.

Also watch how Honor expands the AI features tied to the gimbal, such as gestures and expression recognition, and whether buyers use them beyond filmmaking and vlogging, given its high price.

Nikita Bier is a consumer app builder known for the apps tbh and Gas, which were acquired by Facebook and Discord, and he has led product at X. A moat is a lasting advantage that keeps competitors from taking a company's customers. The argument that AI coding tools erode software moats has been building all year, including in a Forbes column in August.

Early on October 12, Nikita Bier posted that technical moats are mostly over at this point. If you are building something meant to be durable, he wrote, the only moats now are proprietary data, network effects and a physical component.

Bier did not elaborate in the post, which drew more than a million views. His three categories line up with a common argument this year: if coding agents can rebuild most software features quickly, then advantages based on hard engineering shrink, while data others cannot get, products that grow more valuable as more people use them, and businesses tied to hardware, logistics or other physical operations remain hard to copy.

A Forbes column by Boaz Sobrado in August made a similar case, arguing that the only startup moats left in the AI era are data and distribution. Bier's list leaves out distribution and brand, which many investors still count as advantages, and adds a physical component.

The claim is an opinion, not a measured finding. Companies such as frontier model labs and chip designers still compete heavily on technical depth, and it is not clear Bier meant his point to cover them rather than typical software startups.

Bier has long focused on consumer social apps, where network effects decide most winners.

Watch whether founders and venture investors shift toward startups built on unique data, network effects or physical operations, and whether pure software startups find it harder to raise money.

Also watch how incumbent software companies respond as AI coding agents get better, since if Bier is right, more of them will try to defend themselves through data access and ecosystem lock in rather than features.