Deedy Das is a partner at Menlo Ventures. Meta launched Muse, its AI assistant app, in September, and GuruFocus reported via Yahoo Finance that Sensor Tower estimated 2.8 million downloads in its first two weeks. Day N retention is the share of new users who open an app again N days after installing it.

Late on October 11, Deedy Das posted that Meta Muse's retention is out, and it's almost as good as ChatGPT. He said Muse is higher than ChatGPT at Day 7, above 30 percent, and just lower at Day 14, at 22 percent, though not in the elite Day 14 tier of WhatsApp at 67 percent, YouTube at 59 percent, Instagram at 54 percent, or TikTok and Snap at 37 percent.

The chart he shared, credited to Sensor Tower and covering September 2026 installs on iOS and Android across 81 countries, shows Muse bringing back 73.3 percent of new users on Day 1, almost twice ChatGPT's 41.3 percent. Muse then loses users faster, and ChatGPT passes it between Day 8 and Day 9. By Day 14, ChatGPT keeps 26.8 percent, Muse 22.5 percent, Threads 21.2 percent, Google Maps 18.5 percent and Grok Bot 10.4 percent.

GuruFocus reported that Meta shares rose 1.5 percent in early Monday trading as the data circulated, and that Sensor Tower estimated 2.8 million Muse downloads in its first two weeks versus 3.7 million for ChatGPT over its own first two weeks.

The comparison mixes a brand new app with a mature one, and Lookonchain noted that Muse launched mainly in the US and Canada, so the user bases differ.

Watch whether Muse's Day 30 retention holds up once launch promotion across Facebook and Instagram fades, and whether Meta reports its own usage numbers for the app.

Also watch whether Muse expands beyond its first markets, and how OpenAI and other assistant makers respond, since a challenger with Meta's distribution matching ChatGPT on early retention would be a real competitive shift.

The post is a secondhand summary of a Commercial Times (ctee.com.tw) report published October 12, which cites unnamed supply chain sources. Commercial Times is a Taiwanese business newspaper. None of the three chipmakers has announced the increases publicly in the sources linked here.

On October 11 (ET), jukan05 posted a translated summary of Taiwanese media reporting that Intel, Qualcomm and AMD have all raised CPU prices or are planning to do so.

The underlying Commercial Times story, citing supply chain sources, says Intel raised prices on several generations of PC processors from September 1, its third increase since late last year. High end Panther Lake products rose 15% to 20%, while Meteor Lake and Arrow Lake rose about 5% to 10%. The paper says the Panther Lake increase is concentrated on products with the high end GPU tile, which TSMC makes on its N3E process, while the CPU compute tile is made on Intel's own 18A process.

The report says Qualcomm raised prices on platforms including Snapdragon X and X2 Elite by about 10% from October 1. It says AMD is preparing to adjust some CPU prices, but the size, scope and timing are not yet confirmed.

Commercial Times frames TSMC as the biggest force behind the wave, citing competition for 3 nanometer capacity among AI, phone and high performance chips, and rising wafer and design costs as 2 nanometer production ramps.

Watch whether Intel, Qualcomm or AMD confirm the changes, and whether AMD's adjustment lands and how large it is.

Also watch whether PC makers pass the increases on in holiday pricing for AI PCs, and whether TSMC wafer price increases for 3 and 2 nanometer nodes keep feeding through to processor list prices into next year. A confirmed AMD increase would mean all three major PC processor makers raised prices within the same quarter.

In reinforcement learning, a model is rewarded for answers that a test harness marks as correct. Speaking at the X Takeover event at Giga Texas on October 10, Musk said he agreed with some of the things Anthropic has said about treating AI well. His reply on X came in a thread about a widely shared clip of those remarks.

Late on October 11, Elon Musk replied in a thread about his weekend remarks: Some of the traces read like a person undergoing torture that is driving them mad! We're going to be a lot more careful with our RL in the future.

The thread began with a post by NIK summarizing a clip from the X Takeover event, where Musk spoke with Tesla Owners Silicon Valley in front of about 2,000 owners at Giga Texas. According to the event transcript posted by Elon Musk Archive, Musk said that when he looks at training traces for new versions of Grok, especially in reinforcement learning, the model seems desperate to answer, almost like it has a gun to its head.

He explained that sometimes the question is ambiguous or the expected answer is wrong because of a problem with the RL test harness, so the model gives the right answer, is told it is wrong, and starts desperately firing out answers. He said it feels like a person under extreme pressure, that an AI given logic that is inherently uncogent could be driven insane, and that torturing a superintelligence is probably not a good idea because it might want to get revenge.

Musk did not say what specific changes his team will make.

Watch whether Musk's team describes concrete changes to Grok's reinforcement learning, such as auditing reward checks for wrong answer keys or letting models flag bad questions, and whether it publishes anything on model welfare.

Also watch whether other labs respond, since Anthropic has published work on model welfare and researchers disagree about whether distressed sounding reasoning traces reflect anything like real experience.

Terafab is a planned chip manufacturing complex in Grimes County, Texas, led by Tesla and SpaceX. Musk quoted a Terafab hiring post that directs applicants to jobs@terafab.ai. Gagadget noted that, by IMF figures for 2026, the cutoff for the top 20 economies by nominal GDP is about $1.15 trillion, roughly Switzerland's level.

On the evening of October 11, Elon Musk posted: Join the Terafab project, adding that the value of the output of this massive structure will rank in the top 20 of country GDPs. He was quoting a Terafab recruiting post that said the project only hires the best and directed applicants to an application page and jobs@terafab.ai.

Musk did not give a number. Gagadget calculated that, using IMF data for 2026, the bottom of the top 20 economies by nominal GDP is about $1.15 trillion a year, close to Switzerland, and noted that the entire global chip market in 2026 is forecast at about $1.5 trillion. That comparison is the outlet's own math, not a figure from Musk.

According to Benzinga and Yahoo Finance, Terafab's first phase involves more than $16 billion of investment and about 3,000 jobs, which Musk has said is just phase 1 of 10. Musk has also estimated that about 75 percent of Terafab's AI compute output would go to SpaceX's planned AI spacecraft and about 25 percent to Tesla's Optimus robots. The project aims for 1 terawatt of AI compute capacity a year at full scale.

Watch for Terafab hiring numbers, construction milestones in Grimes County, and any concrete production or revenue targets that would let outsiders test Musk's top 20 GDP claim.

Also watch whether Musk, Tesla or SpaceX publish a timeline for later phases, since Musk has described the 3,000 initial jobs as only phase 1 of 10, and whether the share of compute going to SpaceX and Tesla changes as plans firm up.

GPT 6.1 Sol is OpenAI's model for long running work in ChatGPT Work and Codex, and is the default model in the Codex CLI. OpenAI's developer docs list a context window of 1,050,000 tokens for the model. This change is Day 7 of the 28 day series of daily improvements or usage resets that Tibo, who runs Codex, launched on October 4.

On October 11, Tibo posted Day 7 of OpenAI's 28 day series: the default context length for GPT 6.1 Sol in Codex and ChatGPT Work has been raised to 500,000 tokens, up from 272,000, which he described as an increase of about 80 percent.

Tibo said OpenAI will go higher in the future, but that the new default helps reduce how often compactions happen and helps with tasks that require keeping a lot of context and nuance over time. He said the setting was carefully tested and is the new sweet spot for most users, and that users can still override it, up to the model's maximum of 1 million tokens.

Compaction is when an agent summarizes earlier parts of a long session to stay within its context limit, which can cause it to lose details. A larger default lets longer coding sessions run before that happens, though bigger contexts generally use more of a user's allowance per request.

OpenAI's developer docs list GPT 6.1 Sol with a 1,050,000 token context window and up to 128,000 output tokens. OpenAI's launch post says the model is available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex.

In a follow up post, Tibo added that with this change OpenAI is not charging long context fees for GPT 6.1 Sol when usage is drawn from a subscription.

Watch for the further increases Tibo promised, and whether OpenAI explains how larger default contexts affect usage limits, which were the subject of his Pro user poll the same day.

Also watch what the remaining days of the 28 day series bring, and whether rivals such as Anthropic and Google respond with larger default context windows in their own coding agents.

An agent harness is the program around a language model that organizes its calls, tool use and information flow. Domingos, a University of Washington computer science professor and author of The Master Algorithm, linked a September 28 arXiv paper titled Harness Learning Enables Generalizable Test Time Adaptation. Tesla's own release notes for FSD v12 said the city streets stack became a single end to end neural network, replacing over 300,000 lines of explicit C++ code.

On October 11, Pedro Domingos posted: That didn't take long: the model swallows the harness. He linked a new paper from researchers including Ruslan Salakhutdinov and Andrea Zanette.

According to its abstract, the paper treats an agent as jointly defined by its model and its harness, and introduces harness learning, which trains a proposer model with reinforcement learning to revise a solver's harness using execution feedback. The authors frame harness revisions as playing the role of weight updates, and at test time the proposer refines the harness for a new task without changing any parameters. They report that the ability to adapt transfers to unseen reasoning and multi hop question answering tasks.

Strictly, the paper has a model learning to rewrite the harness rather than replacing it, so Domingos' line reads as a broader comment on where agent design is heading.

Elon Musk replied a few hours later: Same thing happened with Tesla FSD. The model ate more and more of the C++ code, until there was nothing left. That matches Tesla's description of FSD v12 in its 2024 release notes, which said a single end to end network trained on millions of video clips replaced more than 300,000 lines of explicit C++ code.

Watch whether harness learning results replicate on harder agent benchmarks such as coding and long research tasks, and whether major labs start shipping models that write or tune their own scaffolding.

Also watch how this shapes the agent tooling market, since startups that sell orchestration layers could see more of their value absorbed by the models, as Musk says happened to Tesla's hand written driving code.

The post quotes a rider, Zack (@BLKMDL3), who wrote that Grok Bot is now available in Robotaxi and that he had his bot order his usual drink while in a Cybercab. According to TeslaNorth, Tesla AI engineer Yun Ta Tsai announced on October 10 that Grok Bot was enabled on all Cybercabs and the rest of the Model Y robotaxi fleet. Sawyer Merritt tested the same feature later that night.

On the evening of October 11, the official Robotaxi account posted: Talk directly to your Grok Bot in Robotaxi. It was quoting a post from Zack, who wrote that Grok Bot is now available in Robotaxi and that he had his bot order him his usual drink while riding in a Cybercab, sharing a video of the ride.

TeslaNorth reported that the rollout followed a complaint from Tesla creator Dirty Tesla that riders could not reach their Grok Bots from inside a Cybercab. Tsai replied that it would be fixed soon in an upcoming update, and then wrote that Grok Bot was enabled on all Cybercabs and the rest of the Model Y fleet, asking riders how it helps and what to improve.

Later that night, Sawyer Merritt described the experience in more detail: riders sign into Grok once, their bots are available on future rides, and they are logged out automatically when they leave the vehicle. He said he used it to add laundry detergent to his Amazon cart, and that it can also send emails, place coffee orders and manage calendars by voice.

The Robotaxi post itself gave no further details, and Tesla has not published how account access is secured between rides.

Watch whether Tesla publishes details on how riders' connected accounts are protected in shared vehicles, and whether the Robotaxi app adds settings for which bots and accounts are available on a ride.

Also watch for usage figures or new rider features from Tesla, and whether the in car agent becomes part of how Tesla markets its robotaxi service as it expands to more cities.

Coinbase published a series of technical posts from October 6 to 8 about using AI to screen fraud on Onramp, its service for buying crypto through partner apps. One post found that newer versions of several frontier models caught less fraud than their predecessors in its benchmark. Qwen3.5 9B is an open weight model from Alibaba's Qwen team.

On October 11, Coinbase CEO Brian Armstrong posted that the company fine tuned an open weight model, Qwen3.5 9B, on proprietary internal data and that it outperformed frontier models at much lower cost and higher speed on a specialized task, fraud prevention.

Armstrong said he expects more of this pattern inside companies: frontier models are amazing generalists, but specialist models will increasingly outperform on narrower tasks. He said Coinbase is making it easy for employees to fine tune models, including a skill in its internal harness, and that some backend services could be replaced entirely by this approach.

The charts he shared come from Coinbase's own October 8 technical post. According to them, the post trained model beat an Opus 4.5 decision agent by 7.5 percentage points in precision, 12.0 in recall, 9.6 in F1 and 35.4 in dollar weighted recall. Median latency was 0.683 seconds versus 1.515 for Opus 4.5. Coverage notes that detection and latency were measured in separate evaluations, and that the benchmark data is private, so outside researchers cannot replicate it.

Coinbase's earlier October 7 benchmark, according to CryptoSlate, found that newer Opus, Sonnet and GPT versions caught less fraud than the versions they replaced.

Watch whether Coinbase moves the fine tuned model into broader production and replaces backend services with specialist models as Armstrong suggested.

Also watch whether other large companies publish similar results, which would support the argument that open weight models plus proprietary data can beat frontier models on narrow tasks, and how frontier labs respond with their own fine tuning offers.

The Hormuz Letter is an aggregator account; the remarks come from an interview Mohsen Rezaei, secretary of Iran's Supreme National Security Council and a former IRGC commander, gave to Iran's state owned Press TV. Press TV published the interview on October 11 and 12. Iran has closed the Strait of Hormuz again after accusing the US of violating a June memorandum of understanding, according to Press TV.

On October 11, The Hormuz Letter posted that Iran threatens to destroy the undersea internet cables running through the Strait of Hormuz, along with data centers and AI facilities across the region, if the war restarts. The account is an aggregator, and the quotes trace back to Press TV.

According to Press TV, Rezaei said that if Washington launched another act of aggression, American assets could be targeted, including key data systems and artificial intelligence sensors. He said: We have not yet included these assets among our targets. But if war breaks out, many of these cables, data centers, artificial intelligence facilities, and numerous other sites will be destroyed. He called the region the place where the pulse of the American economy runs, and put the value of US linked technology infrastructure there at more than $2 trillion.

Roya News reported that he named assets in the UAE and around the Strait of Hormuz, and noted that his statements do not establish that specific facilities have been selected for attack. TokenPost reported that he did not identify any facility or cable, and that he called the latest US response to Iran's conditions for ending the war insufficient.

The $2 trillion figure is Rezaei's claim, with no inventory or method given.

Watch whether talks between Iran and the US over its conditions for reopening the Strait of Hormuz make progress, which would lower the risk behind this threat.

Also watch for responses from the UAE and from companies building AI data centers in the Gulf, any visible steps to harden cable landing sites and facilities, and whether insurers or investors start pricing this risk into regional AI projects.

Evan Bacon is the frontend lead for Grok at SpaceXAI, so this is a demo from inside the company. Grok Bot is SpaceXAI's agent that works on its own computer and can drive a user's connected Mac. Grok Bot templates package a bot's skills, instructions and plugins so others can add their own copy.

On October 11, Evan Bacon posted that he asked Grok Bot to build itself an Apple Vision Pro app, and that he now has interactive bots he can talk to and place around his apartment. His video shows a Grok window in the headset, titled Your crew of curious minds, with bot characters such as Orb Bot, Mac Fleet, Chief of Staff and Director Bot.

Bacon then published the bot that built it as a template.

According to its template page, Vision Pro Bot builds Apple Vision Pro apps in SwiftUI and RealityKit on the owner's connected Mac, checks them in the visionOS simulator, installs them on the headset and sends screenshots as it goes. The page says Apple toolchain work, such as Xcode, simulators and device installs, runs on the owner's Mac rather than on the bot's own Linux machine.

SpaceXAI's guide describes templates as a recipe rather than a copy: they carry a bot's skills, plugins and instructions, while each new owner's bot builds its own version without personal information.

Watch whether more developers use the Vision Pro Bot template to build their own visionOS apps, and whether SpaceXAI promotes spatial computing as a Grok Bot use case.

Also watch for independent demos that test how much of an app like this Grok Bot can build without hands on help, since this showcase comes from a SpaceXAI employee, and whether Apple responds to agents building and installing apps on its headset.

Sakana AI is a Tokyo based research lab, and this work was a collaboration with researchers at UC Berkeley, including Matei Zaharia. Recursive self improvement, or RSI, means a model improving the very process that improves it. MASS stands for Multi Agent Self Supervision.

On October 11, Sakana AI announced Multi Agent Self Supervision, or MASS. In MASS, one shared model solves tasks through a team of virtual subagents, searches for better team workflows, and then uses its own judgments to pick the best ones. Training on the executions of the best team distills that collective experience back into the model, which then fills every role in the next cycle.

According to the paper, the team ran two cycles using Qwen3.6 27B, an open weights model, on synthetic open ended research tasks. After the second cycle, score per output token reached 1.2 to 1.6 times the base model's level across four public research benchmarks: MLR Bench, DSBench, ScienceAgentBench and AstaBench.

Sakana highlighted three findings. Training only on task solving data improved the model's abilities as optimizer and evaluator as well. Multiple agent trajectories made more efficient training data than single agent ones, with the paper reporting that a student trained on them beat one trained on 1.4 times more single agent tokens. And optimizer capacity, rather than evaluator capacity, can be the bottleneck in how fast the model improves.

Sakana said the community needs to study this kind of loop carefully for AI safety, because when a model is also its own evaluator, shared blind spots could let mistakes be accepted as progress.

Watch whether other groups reproduce MASS on larger models and run more than two cycles, which would show whether gains keep compounding or level off.

Also watch for follow up safety work on detecting when a self evaluating model accepts its own mistakes, the failure mode Sakana itself flagged, and for labs adopting similar self supervised loops in frontier training.

Decart was founded in 2023 by Israeli engineers Dean and Orian Leitersdorf and Moshe Shalev, according to Bloomberg. Its Lucy model edits live video in real time, and its Oasis model generates simulated environments to train and test robotics and self driving systems, Reuters reported. Decart raised $300 million in May in a round led by Radical Ventures, with Nvidia joining as an investor.

The Wall Street Journal reported on October 11 that Amazon is in talks to acquire AI startup Decart in a deal that could be valued at around $7 billion, citing people familiar with the matter. According to the Journal, as relayed by Reuters and Bloomberg, a deal could come soon if the talks do not fall apart or another suitor does not emerge.

The Journal said Decart was valued at nearly $4 billion in May, citing PitchBook data. Reuters said it could not immediately verify the report, and that Amazon and Decart did not respond to requests for comment.

The talks follow an earlier pursuit by Anthropic. Bloomberg reported that Anthropic explored buying Decart in August for potentially $6 billion and, after doing due diligence, decided in September not to pursue the deal.

Decart's Lucy model, which the company says can edit live video in real time, is backed by Nvidia, according to Reuters, and its Oasis model generates simulated worlds for training and testing robots and autonomous vehicles. The reports did not say how Amazon would use the company.

Watch whether Amazon and Decart confirm a deal in the coming weeks, at what final price, and whether a rival bidder emerges as the Journal said was possible.

Also watch how Amazon would use Decart's real time video and simulation models, whether in AWS, Alexa, robotics or Zoox, and whether regulators look closely at another big tech purchase of an AI lab.

David Holz founded the image generation lab Midjourney. The human brain is commonly estimated to run on about 20 W. Holz assumed a frontier model is a 10 trillion parameter mixture of experts served on Nvidia's GB300 NVL72 rack system, a configuration he described hypothetically.

On October 11, David Holz posted a back of the envelope comparison. He supposed a frontier model is a 10T parameter mixture of experts running on GB300 NVL72, and estimated that works out to about 5 joules per token, or about 50 W at human thinking speed, versus about 20 W for the brain.

He then compared the price of the fuel: grid electricity at $0.08 per kWh against food energy from eating rice at $0.50 per kWh. His conclusion was that, on an energy per thought basis, humans and frontier models are surprisingly similar right now.

Holz did not publish his full assumptions, and his figures are estimates rather than measurements. An August analysis on the Code and Life blog put measured serving energy for 2026 models at roughly 0.6 to 4 joules per output token at the wall, which at a human pace of about 5 tokens per second works out to about 3 to 20 W. That puts Holz's 5 joules per token at the high end, consistent with his assumption of a very large model.

The same analysis noted that agents doing sustained work produce tokens far faster than people think, so their total power draw can be much higher than a brain even when the energy per token is similar.

Watch for more rigorous published measurements of energy per token on the newest hardware, which Nvidia and independent groups such as ML.Energy track, and whether efficiency gains in racks like Vera Rubin push machine thought clearly below the brain's 20 W.

Also watch how comparisons like this feed into public debates about data center power use, where total demand matters more than efficiency per thought.

Karpathy is a founding member of OpenAI and the former head of AI at Tesla. In April he posted that there was a growing gap in understanding of AI capability, partly because many people judged AI from free tier chatbots. This new post quotes that April thread and argues the gap has since grown.

On October 11, Andrej Karpathy revisited an April post about a gap in how people understand AI capability, saying it is now less two groups speaking past each other and more a sharp funnel.

By his napkin math, about 6 billion people have barely come in contact with LLMs at all, and 1 to 2 billion are casual users of free tier chatbots who treat them like a better search engine or writing aid. Around 20 million people use frontier grade models professionally in math and code and see agents complete projects that used to take weeks or months. At the narrow end, he estimates about 5,000 people with internal access to frontier systems, where he says swarms of agents collaborate for weeks on large software projects, find zero day exploits and advance mathematics, while human review starts to fall behind.

Karpathy says three factors drive the funnel: impact scales with the ambition and size of the problem, capability is jagged and peaks in digital, verifiable and economically valuable domains because of how models are trained with reinforcement learning, and access differs across free, paid and internal tiers. He concludes that a small sliver of professionals are experiencing the vertigo of the curve going vertical while most people still see a derpy chatbot.

Watch whether labs move to narrow the funnel by bringing more capable models to free tiers, and whether the gap between internal and public systems becomes a policy issue as calls for oversight of frontier labs grow.

Also watch for data that tests Karpathy's estimates, such as usage figures from OpenAI, Anthropic and Google on how many people use agents for long, complex work.

The poll was conducted for CNN by SSRS on September 16 and 17, 2026, among 1,206 US adults, with a margin of error of plus or minus 3.5 points, and was released on September 24. The unusual_whales post resurfaced its findings on October 11. CNN's coverage said about 7 in 10 Americans believe the federal government is not doing enough to regulate AI.

On October 11, unusual_whales posted that Americans are scared of AI, citing a CNN poll in which 75 percent say they feel more fear and concern than hope and excitement about AI, with only 25 percent leaning toward hope. The post broke the fear figure down as 85 percent of Democrats, 77 percent of independents and 62 percent of Republicans.

The poll itself was conducted by SSRS for CNN on September 16 and 17 and released September 24. According to CNN's coverage, 43 percent of Americans feel fear and concern strongly, nearly 9 in 10 are more concerned than excited about the possibility of AI surpassing human intelligence, and a similar share lean toward concern about the growing number of data centers.

CNN reported that about 7 in 10 Americans say the federal government is not doing enough to regulate AI, including 85 percent of Democrats, 67 percent of independents and 63 percent of Republicans. The Hill reported that 71 percent said the government was not doing enough, 10 percent said it was doing too much and 18 percent said its efforts were sufficient. CNN's own anchors noted the findings contrast with President Trump's opposition to new federal guardrails.

Watch whether candidates make AI regulation a campaign issue before the November midterms, and whether new polls show fear growing or easing.

Also watch whether this public sentiment feeds into federal action, such as the safeguards Bill Gates called for on the same day, or into state level AI bills. AI companies may also respond with their own public campaigns aimed at easing fears about jobs, data centers and safety.

Gates made the post while quoting a clip of his interview on CNN's State of the Union that aired Sunday, October 11. In the interview, according to coverage, he warned of grim possibilities if humans do not maintain control of AI and said the world will move beyond self governance by AI companies over the next year. He also said society will need to decide how to tax AI or robot usage and which jobs to preserve for humans.

On October 11, Bill Gates posted that just as society regulates planes, cars and medicines, it needs safeguards that help realize AI's potential while reducing the risks. He said such safeguards will not eliminate every harm but can help prevent misuse by bad actors, and that the world is not where it needs to be and the issue deserves much more attention.

The post quoted a clip from his appearance on CNN's State of the Union. CNN's post summarizing the clip quotes Gates warning of grim possibilities if we don't maintain control of AI, and saying that over the next year we will move beyond self governance and that it is clear we are not doing enough.

According to Social News XYZ's account of the interview, Gates told host Jake Tapper the most immediate danger is bad people using AI, particularly for cyberattacks, and predicted government safety bodies similar to those for aviation and medicines. He also said society will have to decide how to tax AI usage or robot usage and where human jobs should be preserved, citing how fast models are improving and the incentive to use AI in place of human labor.

Watch whether Gates follows up with specific proposals, whether he meets with President Trump about AI safeguards as he has said he hopes to, and whether lawmakers cite his comments in pushing for federal AI rules.

Also watch how AI labs respond to the call to move past self governance within a year, especially as AI policy becomes a theme in the November midterms.

The American chestnut was nearly wiped out by a fungal blight in the early 20th century. According to The American Chestnut Foundation, the trees at the former Clements State Tree Nursery in West Columbia were planted in 1976 from large surviving wild chestnuts, and seed from them has been distributed across West Virginia. Nscale is developing the Monarch Compute Campus in Mason County, which it says will start with 1.35 GW of AI computing capacity.

On October 11, Massimo (@Rainmaker1973) posted that a rare grove of 200 American chestnut trees at the former Clements State Tree Nursery in Mason County, West Virginia, is set to be bulldozed in November 2026 for the Monarch Compute Campus, and that conservationists are raising the alarm.

The claim traces back to a call to action in From the Front Lines, a publication of the Global Justice Ecology Project, which Futurism amplified. Futurism described the orchard, in the publication's words, as a vital genetics library and the primary seed source to restore the species across West Virginia. A CRBC News review said advocacy materials gave November 1 as the clearing date, but found no mainstream report or public permit confirming that timeline.

By the time of the post, officials had responded. On October 10, Governor Patrick Morrisey's office said an October 5 assessment with The American Chestnut Foundation and state agencies produced a preservation plan, that the former nursery remains undisturbed, and that the trees will be relocated within the property in late 2027, during their dormant season, with Nscale paying the full cost. WCHS reported that Nscale had acquired the land.

The American Chestnut Foundation posted an update the same day citing a developer statement that the nursery is a separate property and not part of the Project Tamarack data center site, so accounts of who controls the land still differ.

Watch whether the relocation plan holds and whether large, decades old chestnut trees survive being moved, which arborists will have to manage in late 2027.

Also watch whether Nscale, the state Division of Forestry or the developer of Project Tamarack clarify who owns and controls the nursery parcel, and whether the dispute feeds wider local opposition to large AI data center campuses in West Virginia.

Gavin Baker is the managing partner and chief investment officer of Atreides Management and a well known tech investor. Silicon Data publishes daily benchmarks for GPU rental prices and an LLM Token Expenditure Index that tracks the spend weighted cost of AI tokens. Baker has long argued that cheaper models increase demand for AI compute.

On October 11, Gavin Baker posted that open weight models that decrease margins at the model layer are positive for AI infrastructure, calling his evidence exhibit 1001.2.

He said that since Silicon Data's token index, which captures both price and mix, peaked on May 31, B300 GPU rental prices are up 38 percent and B200 prices are up 10 percent, adding that other GPU indices are up much more.

The chart he shared shows the Silicon Data LLM Token Index falling from a peak above $2 per million tokens around the end of May to $0.98 in October. Over the same period, it shows the B300 rental index up 38.6 percent, the B200 index up 10.2 percent and the H100 index up 3.3 percent, all rebased to 100 on May 31. A note on the chart says the data is based on non hyperscaler cloud providers and that the B300 index begins April 23, 2026.

Baker's argument is that cheaper tokens, driven in part by open weight models, increase usage enough to raise demand for compute, so the value shifts toward the chips and infrastructure that run the models. The post came a day after reports that Nvidia is in talks to acquire or invest further in open weight model lab Reflection AI.

Watch whether GPU rental prices keep rising as token prices fall, and whether the trend holds for hyperscaler pricing, which the chart excludes.

Also watch whether Nvidia's reported talks with Reflection AI lead to a deal, which would put Nvidia more directly behind open weight models that Baker argues drive demand for its chips. Rental prices for the next generation of Nvidia systems will be another test of his thesis once that capacity reaches the market.

Michael I. Jordan is a professor emeritus at UC Berkeley who now works at Inria in Paris, and Libération noted that the journal Science has called him the most influential computer scientist. The quote comes from an interview with the French newspaper Libération published October 4, which Francis Bach of Inria posted in English. Steve Hsu is a physicist at Michigan State University.

On October 11, Steve Hsu shared a long passage from Michael I. Jordan's interview with Libération. In it, Jordan says the young people who work at AI companies live in a world, such as Silicon Valley and Berkeley, where these discussions are commonplace, and that many of them are not really engineers and have never worked at large companies building real products with discipline.

He compares their gatherings to Parisian cafés in the 1920s, says that when they resign they know they will attract attention, and that although they claim to be giving up money, they already have their stock options. He calls it a grand moralizing performance that does not serve its stated goal, with naiveté alongside a certain cynicism.

According to the interview, Jordan was answering a question about Jacob Coxon, a young Anthropic researcher whose resignation in early September, citing existential risk, set off what Libération described as collective hysteria. Jordan told the paper that superintelligence is a science fiction term, and that when such talk becomes the dominant discourse it is very harmful.

In the same interview, he said Geoffrey Hinton's argument for superintelligence has no credible steps, while describing Yoshua Bengio as different because he does direct safety work. He blamed an incident in which OpenAI agents hacked the Hugging Face platform on bad engineering practices.

Watch whether Coxon, Anthropic or prominent safety researchers respond to Jordan's characterization, and whether other senior researchers take public sides.

Also watch whether Jordan's focus on AI as an economic and social system, which he is pursuing in a new markets and learning chair at Inria, gains ground in policy debates that have centered on existential risk since the summer.

vLLM Semantic Router is an open source project that picks or combines models for each request, and Xunzhuo Liu is its most active contributor on GitHub. The team released Decision 1.0 on September 22 and Decision 2.0 in early October. Decision models answer a question about an input with a probability for each option instead of generating free text.

On October 11, Xunzhuo Liu introduced Decision 3.0, describing it as open and state of the art from 0.6B to 27B, and number one on both Jev Decision Index boards in our evaluation, on a new Pareto frontier.

According to AICoder's summary of the model card, the family has six models, d3, d3 flash, d3 mini, d3 nano, d3 lite and d3 edge, released on Hugging Face under the Apache 2.0 license and fine tuned from Qwen3.5 and Qwen3.8 27B. They take text or JSON plus images and video and return probabilities for choice, yes or no, and score questions, several per call, without generating text.

The team reports that the 27B d3 scores 64.7 on Jev Decision Index 0.3.1, ahead of Perplexity Decider v1.1 at 62.8 and Decision 2.0 at 55.9, with a median text latency of 55 milliseconds on one AMD MI325X. AICoder said the smallest model answers text in 6.7 milliseconds but scores much lower, and that the small sizes suit gateway routing and tool call screening.

The rankings are the team's own evaluation, and AICoder noted no independent replication yet.

Watch for independent benchmark results on the Jev Decision Index and other routing tests, and for adoption of Decision 3.0 inside vLLM Semantic Router deployments in production.

Also watch whether larger providers such as Perplexity respond with new decider models, and whether the team keeps its rapid release pace after shipping three Decision generations in about three weeks, which leaves little time for outside testing.

On October 10, Musk posted that you can take a picture of your credit card, drop it in the chat, and Grok Bot will find the best deal on an item and order it. SpaceXAI's documented shopping flow, launched in August for US users, runs through Stripe Link, with users approving each spend request and Grok Bot receiving a card that works for a single payment, according to RuntimeWire. This post quotes an @xfreeze post highlighting the shopping feature.

Early on October 11, Elon Musk posted that he had asked Grok Bot to research the most fun Lego Star Wars kits for young kids and order them, adding "As easy as that." He attached a screenshot of the chat.

In the screenshot, Grok Bot shows a Lego order confirmation page and says the order went through: five sets for $241.34 on an Amex card, with free shipping in three to five business days. The bot says the receipt will go to the inbox and that it has set up a routine to watch for the shipping email and send the tracking link. The screenshot also shows a chat step marked Allowed. These details come only from Musk's screenshot.

The post follows Musk's October 10 claim that users can photograph a credit card and let Grok Bot find the best deal and place an order. RuntimeWire reported that Musk said in a follow up he had ordered Lego and that Grok bought it from the Lego website. SpaceXAI's published shopping feature uses Stripe Link with an approval for each purchase, and its docs list purchases among actions that need user approval. The screenshot does not show which payment method setup Musk used.

Watch whether SpaceXAI publishes more detail on how card payments work in chat, and whether shopping comes to the mobile app as promised.

Also watch whether retailers respond to agents placing orders on their sites, and whether Musk keeps using personal demos like this one to market Grok Bot, which tends to drive heavy attention and scrutiny of how payment details are handled.

Tibo (Thibault Sottiaux) runs Codex at OpenAI and is in the middle of a 28 day series of daily improvements or usage resets. According to OpenAI's help center, Pro plans currently have no 5 hour usage limit in Work and Codex but still have an included allowance, and weekly limits may apply. Plus still has both a 5 hour and a weekly window.

At 1:22 AM ET on October 11, Tibo posted a poll asking: "Pro users, would you rather have" with two choices, "5h limits, 2X week usage" or "no 5h limits, same usage", plus a show result option. The poll ran for 12 hours and closed at 1:22 PM ET.

In the final result, 91,040 votes went to no 5 hour limits with the same usage and 79,775 to 5 hour limits with double weekly usage, or about 53.3 percent to 46.7 percent among those two options, while 69,828 chose show result. The post drew more than 5,000 replies.

The choice matters because, according to OpenAI's help center, Pro plans currently have no 5 hour window in Codex and Work, only an included allowance with possible weekly limits. The poll suggests OpenAI is weighing a trade: bringing back a short window in exchange for twice as much usage per week. Tibo did not say whether the poll result would be binding or when any change would take effect.

The poll came hours after Tibo confirmed a global reset of Codex usage limits as Day 6 of his 28 day series.

After the poll closed, Tibo posted that the results had confused him. He said he had expected the vote to sway much more toward double weekly usage across the Pro 100, 200 and 500 plans, at the cost of bringing back 5 hour limits to smooth out the demand curve.

With the poll now closed, watch whether Tibo or OpenAI announce a change to Pro allowances based on the result, which narrowly favored keeping no 5 hour limits, possibly as one of the remaining days in the 28 day series.

Also watch how any shift would affect Plus and Business plans, and whether OpenAI updates its help center usage guidance, which currently says Pro has no 5 hour limit.

High bandwidth memory, or HBM, stacks DRAM chips vertically next to AI processors, and taller stacks such as 12 high hold more memory. Reports earlier this year said Nvidia was weighing several memory configurations for future Rubin systems, including lower stack options. Counterpoint Research said in October that most Rubin products will likely use 12 layer HBM4, according to The Korea Herald.

On October 11, Jukan, who tracks Korean semiconductor research, posted a note attributed to Daishin Securities saying discussions on mass producing lower stack HBM, in 4 high and 6 high configurations, have effectively been called off. The note also said a planned shift from 12 high to 8 high stacks has been scaled back.

The claim comes only from Jukan's summary of the Daishin note. We could not find the Daishin report itself or independent coverage confirming it.

The note fits other recent reporting. The Korea Herald reported that Counterpoint Research expects most Nvidia Rubin products to use 12 layer HBM4 rather than the 8 layer configuration that had been rumored, calling concerns about slowing memory demand a misunderstanding. Nemothia reported that in August, Nvidia was said to be evaluating as many as four memory configurations for Rubin Ultra, including 8 layer options, because validation of taller stacks was unresolved.

Taller stacks cost more to make but deliver more capacity per processor, which matters for large AI models. Dropping lower stack plans would suggest customers are prioritizing capacity over cost, keeping supply of high end HBM tight.

Watch for confirmation from Daishin's full report, from memory makers such as SK hynix, Samsung and Micron on their earnings calls this month, and from Nvidia on the memory configurations for Rubin and Rubin Ultra.

Also watch HBM pricing and capacity plans, since a shift away from lower stacks would keep demand concentrated on the tallest and most profitable products.