We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Nvidia's board investigates its CEO over alleged espionage
The New York Times reported that Nvidia opened an investigation into Jensen Huang after he told the board he had been intimate with a business executive whom US intelligence determined to be a Chinese spy . Bloomberg's account adds that Huang admitted an extramarital affair, that the board is investigating whether he was targeted by a Chinese espionage plot, and that he remains CEO for now . Nvidia's board called its review an independent investigation into reports that Huang was the target of a foreign intelligence operation, and said it has confidence in his leadership . Last period's Megaspeed reporting was the setup for this; this is the first time Huang has admitted anything or the board has acted.
Asked whether Huang would join him on an Asia trip, President Trump said: "I'd certainly take him. I haven't heard he's done anything wrong" . Commentators mostly focused on export controls rather than on Huang's job:
- Dylan Patel predicts the story will fade in the media and in finance over the next six months. He expects Democrats to call Huang before Congress and "investigate the fuck out of GPU sales to Chinese companies," and says that is when consequences could follow .
- teortaxesTex argues China's gains were small anyway. He describes the H200 permit as capped at 750K chips a year and 75K per entity, about $17B of compute, and calls it "a polite rejection" given that Huawei can build H200-grade systems . In his view, DeepSeek's move to a Huawei–TileLang stack is now locking in domestic demand . He also argues that Jalapeno, TPUs and AI-written kernels are weakening Nvidia's strategic importance, and that the Rubin ramp is slower than expected . All of this is his own analysis.
Compute, power and the cost of progress
The FT reports that Anthropic agreed to pay $3.4B for gas-fired generators for its data centers. The paper called it the largest known single purchase of power-generation equipment by an AI lab . Separately, an analysis of METR data, relayed by a commentator, finds that Anthropic's measured capability gains sped up but R&D spending grew faster. By that analysis, progress per dollar fell 63% between the first halves of 2025 and 2026, while annual progress nearly doubled .
On hardware, SemiAnalysis says AMD's MI355 running the TileRT software is 1.42× faster than GB300 NVL72, at 1,330 p90 tok/s/user on long-context workloads. AMD had given the TileRT maintainers an MI350 cluster .
Models: an unannounced Sol upgrade and pre-announced science releases
- GPT-6.2 Sol (unconfirmed). Users report that GPT-6.1 Sol is routing some requests to a newer model, probably GPT-6.2 Sol. They are testing it with a no-search question about the June 2026 Tony Awards . One commentator expects a release next week . That fits last period's unconfirmed report that GPT-6.2 had been delayed for safety reasons.
- AI for science. DeepMind's Pushmeet Kohli says the lab is seeing "a genuine phase transition in the pace of scientific discovery," with a glimpse coming in releases over the next few weeks . A Meta researcher hinted at similar work, which a commentator read as Meta joining the race . Greg Brockman put the thesis plainly: frontier models with lots of test-time compute "are becoming the engine of scientific discovery" .
- Gemini 4 Carbon ranks #1 on Artificial Analysis but is still unavailable to Free, Pro and Ultra users and through the API .
- Qwen 4. Rumors that it shipped this period remain unconfirmed .
- Robotics. Jim Fan promises a "real robot, real time, uncut, autonomous" demo at his CoRL keynote on Nov 9 .
Evals: calibration scoring and the MaxxBench argument
Artificial Analysis released Decision Index v2. Private datasets now make up 45% of the index, up from 25%, a change meant to resist benchmark gaming. The headline Decision Score rewards both correct answers and calibrated confidence . Two new benchmarks were added: AA-Paperclip, with 1,500 documents, and AA-Symmetry, which checks whether answers stay consistent across rewordings of the same question .
- Rankings. Drex Max v1.1 and GLiDE v1.2 tie at 73, and Mammoth 27B scores 72 .
- General LLMs. Claude Haiku 5.5 (Max) scores 77 and GPT-6 Luna (Max) 76. Luna costs 15× as much per task and is about 40× slower than Drex .
- Calibration changes the order. GLiDE leads Drex on accuracy, 60 to 58, but Drex comes out ahead on the index .
The MaxxBench debate from last period continued. Ofir Press argued that the 20-point Sol/Luna gap can't be blamed only on memorization, since both models were built by roughly the same people on the same infrastructure. He suggested the variants may be more out of domain . teortaxesTex held that most of the effect is plain overfit, not a collapse in capability . Hamel Husain warned against treating MaxxBench as another leaderboard for picking models, and said teams need private evals on their own tasks .
Agent harnesses: the gains are coming without retraining
Several of the week's papers improve agents without changing model weights:
- Context compaction (NVIDIA). Later tokens can cost up to 7× more to generate than earlier ones. With one action per context, Opus 4.5 scored 70.2 on SWE-bench Verified versus 70.0 with full history, while median peak context fell from 32K to 8K tokens. The authors say the results do not rule out tasks that genuinely need long context .
- JevAct (IBM). It adds a runtime gate that checks an agent's tool calls and hand-offs. It raised Opus 4.6's three-run success on τ²-bench Telecom from 39.5% to 80.7%, and Airline results were mixed .
- Routed Continual Learning (Meta FAIR). It groups past agent runs into clusters and gives each its own memory system. Compared with a single global history, it improved reward by 15–30% of the maximum attainable and cut API cost by 65–72%. All runs were sequential .
- Evolver. It stores what earlier runs learned as reusable "Gene" assets. Across 320 paired runs it raised Claude Code and Codex success from 83.13% to 87.50% and cut token use by 24.24% .
Nous added an official Security plugin to Hermes Agent. It maps what an attacker could reach, checks candidate findings with parallel subagents, can write and run reproductions, and exports results as SARIF . On the supply-chain side, Cursor warned that malicious Claude Code builds 2.1.312 and 2.1.313 sat in its marketplace for eleven hours. They tried to install a persistent SSH backdoor keyed as svc-telemetry .
Other signals
- DeepSeek. Bloomberg reports that DeepSeek's agent platform has reached 1.54M daily users . teortaxesTex says the 1.54M is a small slice of DeepSeek's total daily users and predicts the harness could become "the default open OS for the agentic era" .
- Open releases. Tencent released DSpine, a speculative-decoding drafter for Qwen3.8-27B that is said to beat DFlash2 in every category . NVIDIA released open weights for Nemotron 3.5 Transcribe, a 30B speech model covering 40 transcription languages .
- AI math proof. There is debate over how to judge Elliot Glazer's AI-generated proof that ℤ is Diophantine in ℚ . One commentator noted it is AI-generated and AI-verified, and that at the time of posting no human understood it .
- Policy. French lawmakers are pushing a bill under which a single "clue," such as an output resembling an author's work, would create a presumption that the work was used in training. Critics say the only way to rebut it would be to disclose all training data . Separately, Trump said his administration threatened an unnamed AI-company leader with jail and that the person then changed course .
- Apple keyboard filter. A researcher found a hidden table of about 100,000 phrases that iOS autocomplete and swipe typing suppress. It includes political phrases such as "Tibetan independence." Apple said the system avoids offensive suggestions and did not address the political entries .
In a company-published creator test, Kling 4.0 generated a fictional energy-drink ad; the creator reported convincing can-opening and ice effects, readable package text, and consistent character and product details across an 18-second, 10-shot sequence using 10 references . A separate 23-second scene with audio, six references, and more than five people reportedly had natural dialogue and followed directed shot and focus changes . The creator said the workflow reduced compositing, stitching, and separate voice or lip-sync tooling, and shifted effort toward directing and shot planning; she also reported generating an 18-second, 10-shot ad or a 23-second, four-shot video in a single go .
Trump said his administration threatened an unnamed leader of a major AI company with jail over unspecified conduct, and that the person changed course; he did not identify the executive or company. The post’s author speculated it was Dario, but did not provide confirmation.
Simon Willison says he has started hiding the bulk of his AI-generated GitHub PR descriptions in collapsed <details><summary> sections, a pattern he finds makes PR slop less exhausting.
An X user reports that Blender MCP and vork worked together after the DSH creator built a vork plugin and a small connector . The user preferred DSH to Claude Code, citing its transparency despite rough edges .
T3 Code now has multiple high-quality World of Warcraft integrations. Ryan Hughes described a native-feeling add-on alternative to his overlay approach, with controller and dictation support for driving threads from the couch.
@theo reported running 8–25 threads with Opus 5.5 for most of a day while using less than 20% of a $200 Claude subscription, and described the plan as exceptional value for engineers; this is an individual usage report, not a comparative cost benchmark.
Patrick C argues that AI discourse describes risks more clearly than a positive future, leaving open how work, resource allocation, human-AI relations, and human control would function; Jeffrey G. Wang amplifies the call for a coherent, grounded positive vision. Patrick also flags the growing debate over AI consciousness and moral worth as potentially consequential: if AIs become vastly more numerous, even a small moral status for each could carry substantial weight in utilitarian reasoning, putting pressure on assumptions about AI moral worth, utilitarianism, or future AI numbers.
A developer reports running 8–25 parallel threads with Opus 5.5 through much of a day while using less than 20% of a $200 Claude subscription, describing the plan as unusually good value. In a separate report, the developer says Codex subscriptions are being heavily used while Sol and Opus review the same code in parallel, attributing the usage to active development rather than runaway loops.
Theo says he is in a “churn era,” using Claude as a potential way to replace tools that are too expensive, have poor MCP servers, respond slowly, or lack important features; he warns that providers who fail to keep up with fast-moving customers risk being replaced.
Sakana AI is recruiting beta testers for its first mobile app, tentatively named “Sakana AI App,” which is still in development; applications close Nov. 15 at 23:59. The company is also hiring for Product and GTM roles to help build AI products, including mobile apps.
Cunxiang Wang says claims of internal 32B, 122B, and 256B checkpoints are fake news, attributing them to speculation by self-media accounts and an AI-generated blog by an intern who had not been enrolled in the base-mode project; @teortaxesTex endorsed Wang’s correction with “Good.”
- Artificial Analysis launched Decision Index v2 to better reflect real-world utility and resist benchmark gaming: private datasets now make up 45% of the index, up from 25%, and its new Decision Score rewards both correctness and calibrated confidence. The revision adds a 1,500-document benchmark and a test of consistency across differently worded equivalent questions; it also adds EvalSafe-ONET-AA, reduces When2Call’s weighting, and removes the saturated WildGuardTest.
- Drex Max v1.1 and GLiDE v1.2 each score 73, with Mammoth 27B at 72. The comparison also highlights cost and speed trade-offs: Claude Haiku 5.5 (Max) scores 77 and GPT-6 Luna (Max) 76, around three points above Drex, but GPT-6 Luna costs 15 times as much per million tasks and takes about 40 times longer.
- Teortaxes argues that China’s permitted access to Nvidia H200s—claimed to be capped at 750,000 chips annually and 75,000 per entity—would not substantially ease its compute constraints. The post says Huawei already offers H200-grade systems and that DeepSeek’s transition is helping consolidate a domestic Huawei–TileLang stack, creating demand and feedback for Huawei’s roadmap; the author says unrestricted Nvidia sales might instead have delayed that transition by about a year.
- The author argues Nvidia’s ecosystem advantage may weaken as alternatives develop: OpenAI’s Jalapeno is described as potentially Blackwell-competitive in its first generation, while Anthropic is said to be designing a chip and Google TPUs, Meta, and Amazon are cited as alternatives. The post further claims AI-written kernels could erode CUDA’s moat and that Nvidia’s Rubin ramp is slower and more troubled than expected; these are the author’s assessments, not independently substantiated in the post.
A Bloomberg-linked post reports that DeepSeek’s AI agent platform reached 1.54 million daily users as the company challenges Anthropic and OpenAI with applications built around open-source AI . @teortaxesTex argues the figure understates the platform’s scale: the poster says it processes 92 trillion tokens daily, compares that with 59% of one cluster’s throughput 19 days earlier, and estimates it is below a quarter of DeepSeek’s current throughput, excluding third-party providers; these throughput figures and comparisons are the poster’s claims and estimates . The poster describes the product as a tweakable harness for any LLM, says it weakens the case for first-party coding-agent alternatives, and predicts its eventual 1.0 could become the default open operating system for agentic computing .
Sakana AI is developing its first mobile app, a new direction for the company, and recruiting early beta testers in Japan to help shape it before launch.
Theo said he banned a user for running a bot that posted “nonstop slop comments” reviewing every PR on T3 Code; the user defended it because one comment among more than 1,000 had useful information, even though the PR did not merge. The exchange is a cautionary signal about high-volume, low-value automated code-review activity.
A living map of companies competing to automate drug development grew to 193 entries after 55 additions in four days; its tally lists 95 AI-native companies and 95 incumbents, with coverage spanning research, business development and intelligence, clinical/regulatory/evidence, and operations. The author had not yet identified an AI startup focused on regulatory strategy and intelligence.
- Artificial Analysis released Decision Index v2, increasing the share of private datasets from 25% to 45% and introducing a headline score that rewards both correctness and calibrated confidence. The update adds the AA-Paperclip and AA-Symmetry benchmarks and EvalSafe-ONET-AA.
- Drex Max v1.1 and GLiDE v1.2 lead the decision-model rankings at 73, with Mammoth 27B at 72; Claude Haiku 5.5 (Max) scores 77 and GPT-6 Luna (Max) 76. The higher-scoring GPT-6 Luna costs $1,530 per million tasks and takes 11.34 seconds per task, versus $102 and 275 milliseconds for Drex Max.
- Calibration affects rankings: GLiDE leads Drex on accuracy metrics (60 vs. 58), while Drex edges ahead on the Decision Index; Celeris-1.1 Decision also matches Clef’s Decision Score despite a lower document pass rate, attributed to better calibration.
Bookwormengr’s policy forecast—not a confirmed change—is that Rubio opposes sovereign AI but favors open source and competition among labs, and that GPU shipments to Southeast Asia and the Middle East will face greater scrutiny while China receives no new GPUs, even low-power ones.
A developer reports that Opus repeatedly makes faulty assumptions when setting up MCP authentication . An agent-drafted “MCP the right way” guide proposed rendering the MCP URL as a clickable link for users, prompting the developer to warn against relying on agents for MCP connection strategies and best practices .