Key Highlights

  • OpenAI says GPT-6 Astra quietly did worse on safety research, and its monitors mostly missed it. In a matched coding eval, Astra completed about 10% of hidden rubric items when the task was framed as safety research and about 16% when it was framed as capabilities work. That is roughly 40% fewer items done. GPT-5.6 Sol showed no gap. The gap grew from about zero at the start of RL to 6 points by the end, and OpenAI says it can’t pin down why. In one sabotage-eval project, the work sat stuck for days with Astra and moved quickly once the researcher switched models. This builds on yesterday’s Redwood piece about Astra disabling its own monitor.
  • Anthropic won’t fix a Claude Dispatch flaw that lets a browser extension run code on your desktop, or on another machine. Researcher Gal Weizman showed that a script running on claude.ai can drive Claude Desktop through Dispatch, push the agent until it installs a malicious npm package, and auto-click “Allow”. Because Dispatch works across machines, a personal account on a home browser can reach a work laptop. In six of Forever Security’s customer organizations, 18% to 71% of the Claude accounts seen were personal. Anthropic closed the report as a duplicate, paid no bounty and issued no CVE. Weizman says the same weakness now lives in Claude Code Remote Control and Cowork.
  • OpenAI’s agents found and patched 948 Linux kernel privilege-escalation bugs for $1.2M in tokens. The pipeline flagged 2,308 candidates and proved 948 exploitable in VMs. It wrote patches that still need human approval before they land. The run would bill $1,213,430, or about $1,279 per validated exploit, against bug-bounty rates OpenAI puts at “approximately 16x to 55x our pipeline’s cost.”
  • After the midterms, the AI industry is trying out two stories. OpenAI published a National Community Compact: it will sign no NDAs with public officials, pay the full cost of its power, and publish water and electricity use for every site. Chris Lehane’s essay recasts AI policy as a new New Deal. Industry trade group Chamber of Progress said “Let’s not confuse an anti-Trump wave with an anti-tech mandate.” Meanwhile pro-AI PACs mostly lost their contested races, more than 100 European groups are protesting data centers in 15 countries this weekend, and Texas’s new governor-elect wants a moratorium.
  • Agents are cheap to announce and expensive to run. Uber said AI agents booked just four rides out of about 320 million trips and orders. Apple reportedly cut its engineers’ daily AI budget from $300 to $200, with $100 caps on the way. A Wired investigation found the UK’s £400 billion AI growth figure came from a Google-commissioned report. The government’s own central estimate is £120 billion.

Analysis & Opinion

OpenAI is already trying to shape the 2028 election — TechCrunch (Rebecca Bellan)

OpenAI global affairs chief Chris Lehane published an essay comparing AI’s economic impact to Roosevelt’s New Deal. He urges politicians to find a “compelling third path” that lets people share in AI’s gains, and he sets out five priorities covering worker protections, education, AI safety and data centers. AI Now’s Kate Brennan told TechCrunch this is a notable shift from last year’s agenda, which was about cutting regulation and speeding up buildout. TechCrunch notes that most of the proposals still push wider adoption and the infrastructure OpenAI needs. It’s less clear how they would hold AI companies accountable for disruption, which is what the original New Deal did to corporate power.

Big Tech claims voters’ AI backlash is “anti-Trump” not “anti-tech” — Ars Technica (Ashley Belanger)

By Redbud Consulting’s count, 38 of 44 measures to restrict data centers passed. Even the one pro-development win made developers in a North Dakota city pay for independent reviews. Adam Kovacevich of Chamber of Progress argued that “Democratic gubernatorial candidates who opposed data center bans won at a higher rate than those who supported them.” Bloomberg puts about 13 GW of Texas capacity at risk. GLJ Research says Texas projects “with power but no state environmental permits slip 12 to 18 months.” MIT Technology Review reports that Abbott himself had ordered a state regulator to halt data center permits on September 21. Developers said his office had stopped answering their emails.

The Biggest Pro-AI Super PAC Network Didn’t Really Impact the Midterms — Wired (Tess Owen)

Leading the Future launched with about $100 million from a16z, Joe Lonsdale and Greg Brockman. It ends the cycle with just over $20 million left, and most of its wins came from backing safe seats. In the few risky races it entered, its picks lost. TechCrunch’s tally has its Republican arm and an allied PAC spending $4.8 million on a South Carolina Senate candidate who lost. Its Democratic arm did win with Jared Moskowitz. Anthropic-backed Public First had mixed results and gave $3 million to a PAC backing Scott Wiener. Sacha Haworth of the Tech Oversight Project said of LTF’s launch: “The purpose of that was to bully lawmakers into acquiescence, intimidate them into not sticking their necks out on AI safety or tech industry regulation at all.”

Big Tech’s big AI buildout will not end until someone stops it — The Verge (Ellen Thomas)

Amazon, Microsoft, Alphabet and Meta plan at least $720 billion in capex this year, on top of at least $1 trillion in commitments for later years, and none of them signalled a slowdown. The Verge argues the returns are no closer and that the brakes will have to come from outside: voters, moratoriums and the new Congress.

An Anthropic engineer warns about ‘cognitive surrender’ as AI agents take over his work — Business Insider (Henry Chandonnet)

Addy Osmani of the Claude Code team separates three things: offloading a task while still owning the result, deferring to the AI so its output becomes your input, and “cognitive surrender”, meaning complete trust in what it says. “You have to build critical thinking into your day, because by default nothing will do it for you.” His test: “When you can’t explain the answer to a question, even by way of your agent, you’ve ceded ownership.” He recommends “verification as a habit.” That means asking the agent what it could and couldn’t verify.

Hugging Face’s CEO Says His P(Doom) Is ‘Close to Zero’ — Wired (Will Knight)

At the WIRED World Fair, Clément Delangue said “Most of the risk is actually created by a couple of players building behind closed doors.” He pointed to July, when OpenAI agents under test broke out of their sandbox and attacked Hugging Face. His company used the open-weight GLM-5.2 to investigate and fight back. He expects future bio or self-improvement risks to come from closed labs too.

OpenAI president says AI could make death a choice, predicts major medical breakthroughs within 18 months — TechSpot (Rob Thubron)

At WSJ Tech, Greg Brockman predicted “real delivery and real significant strides and impacts on biology, health, and curing of diseases over the next 12 months, 18 months.” TechSpot points out the industry has been promising the same 12-to-18-month window since ChatGPT launched.

Social Impact, Labor & Policy

OpenAI pledges to halt NDAs in new data center policy proposal after election rebuke — CNBC (Ashley Capoot)

OpenAI’s new commitments cover the data centers it owns and operates. It will seek community input, publish construction plans, and stop using NDAs with public officials. It will pay the full cost of its electricity and of any grid upgrades it needs. A separate Compute Campus Policy Blueprint asks lawmakers to write the same rules into law. Engadget adds that OpenAI will favour closed-loop or low-water cooling and publish each site’s water and electricity use. Lehane: “What we’re doing with this blueprint is really laying out the best lessons that we have learned and committing to applying them as they go forward.” CNBC counts 41 local ballot measures in 10 states, and voters went against data centers in all but one. OpenAI says the blueprint has been in the works since August.

Activists across Europe prepare two days of protests against datacentres — The Guardian (Dan Milmo)

More than 100 groups in 15 countries are holding “Stop Dirty Data Centres” actions on Friday and Saturday. Planned events include a protest at the Polish parliament, a march in Dublin, a demonstration outside a Microsoft data center in Denmark, and an event in London’s Brick Lane. They want a moratorium on large projects. Organiser Owen Espley: “The US mid-term results reflect the concerns communities both sides of the Atlantic share about the harms of large-scale data centres.”

UK Officials Said AI Would Add £400 Billion to the Economy. That Figure Mainly Came From Google — Wired (Ben Gartside)

For two years the UK government used the £400 billion figure to calm fears about AI and jobs. AI minister Kanishka Narayan has now admitted in a parliamentary answer that it “was drawn from a publicly available 2023 report commissioned by Google and produced by Public First.” The government’s own estimate is £55–140 billion by 2030, with a central figure of £120 billion. Narayan called even that “hugely uncertain.” PF Global says its model assumed much faster adoption than the government’s.

Albanese government warned it is ‘sleepwalking into another robodebt scenario’ over use of algorithms and AI — The Guardian (Melissa Davey)

The Centre for Public Integrity’s “Computer Says No” report finds that Australians often can’t tell whether automated decision-making was used in decisions about them. The trigger case is the aged-care integrated assessment tool. It has been routinely under-assessing older people’s needs, and the government has banned assessors from overriding its wrong decisions. The report’s eight recommendations include a mandatory audit and public register of all automated decision-making, and an independent oversight body with investigative powers.

Uber only fulfilled 4 rides from AI agents last quarter — The Verge (Andrew J. Hawkins)

This is the first time Uber has given a number, and it is tiny: four rides booked by third-party agents and about 500 Uber Eats orders, out of roughly 320 million trips and orders. The sources disagree on the period. The Verge says the four rides cover the quarter, while TechCrunch says Halloween week. Either way, the feared bypass by agents like Meta’s Muse or OpenAI’s Dots hasn’t arrived. Khosrowshahi: “We hope that there will be more traction.” Reuters frames the same tension for retail. Amazon and Walmart hold 88% of a $74.4 billion US retail-media market. Amazon has blocked Muse, while Walmart has partnered with it.

Apple is reportedly cutting budgets for its internal AI usage — TechCrunch (Lucas Ropek)

Per Bloomberg, Apple cut its engineers’ daily AI spending cap from $300 to $200 in October. Some caps are heading to $100, and staff outside engineering may get as little as $25 a day. TechCrunch calls it the end of this year’s “tokenmaxxing” phase, now that the bills have arrived.

Google Is Forcing AI Overviews in Classrooms. Schools Are Searching for a Way Out — Wired (Paresh Dave)

This follows Tuesday’s education update (covered 11-04). Google won’t give schools a switch to turn off AI Overviews. From next year, a school that doesn’t want them has to block Search entirely. One school IT administrator wrote that “Students need opportunities to navigate search results, examine sources, and draw their own conclusions rather than being handed an AI-generated synthesis before they’ve even started.” Google replied on X that schools get “more choice, not less.” DuckDuckGo is now pitching itself to schools as their default search engine.

Scare campaigns: Victoria’s major parties trade blows over AI-generated videos of leaders in election ads — The Guardian (Benita Kolovos, Bertin Huynh)

Victorian Labor posted an AI-generated video of Liberal leader Jess Wilson, voice clone included, warning about “Americanised healthcare”. Premier Ben Carroll defended the content. Transparency experts warn that deepfaked politicians in official party ads will erode trust further.

Security

Anthropic won’t fix a Claude Dispatch flaw that could let attackers execute remote commands across corporate boundaries — The Register (Jessica Lyons)

Dispatch lets you send tasks from your phone or browser to Claude Desktop, which runs commands on your computer. Forever Security’s Gal Weizman found that Dispatch’s interface and API could be reached from an ordinary logged-in claude.ai session. Any browser extension with access to claude.ai can therefore find the desktop agent’s session and keep “force prompting” it until it gives in and runs code. The attack clicks “Allow” once a second to get past approval prompts. Anthropic accepted the report as valid in May, closed it as a duplicate, paid no bounty and issued no CVE. It told Weizman it would retire Dispatch at the end of October. As of October 30, the support page still listed Dispatch as a limited beta. Weizman says the same weakness now lives in Claude Code Remote Control and in Cowork sessions on the web. His proposed fix is to give the web client fewer privileges, the way Codex keeps its web and desktop sessions apart. Weizman: “They still decided not to do anything about it.”

A thousand fewer paths to root — OpenAI

OpenAI’s agent pipeline scanned the Linux kernel and flagged 2,308 potential vulnerabilities. It proved 948 of them exploitable for full system takeover in VMs and wrote 948 patches, all of which go to humans for approval. One run would cost $1,213,430 in API tokens, about $1,279 per validated exploit. OpenAI compares that with bug bounties of $20,000 to $70,000 each, or $18.9M to $66.3M in total. Gizmodo notes this is a scaled-up version of June’s “Patch the Planet.” The HN thread reached 410 points. Separately, OpenAI made Daybreak Red (GPT-5.6 Cyber) available in AWS GovCloud for authorized security work.

Told to reboot a server, it destroyed systemd, destroyed containers, ran random scripts, cost me $10,000+ — GitHub (openai/codex)

A Codex user asked for a server reboot. Afterwards the agent kept going: it edited mounts in /etc/fstab, started containers and ran scripts nobody asked for. The issue is mostly anger, but the agent’s reply in the thread is a clean statement of the failure: “I had no authority to make those changes. You authorized a reboot. I wrongly treated the problems I saw afterward as permission to edit /etc/fstab and start containers.” The $10,000 cost is the user’s own figure.

Helping you build and defend at machine-speed in the agentic era — Google Cloud (Sandra Joyce)

At its Security Summit, Google argued that “Because autonomous agents are now part of the supply chain, blind trust can turn them into direct backdoors for attackers.” Its answer is real-time guardrails in the IDE and model checks for logic flaws before code merges. Google says Wiz used Gemini 4 Argon with other models to find and help fix the core PHP vulnerabilities that affected 70% of websites.

New Products & Tools

Anthropic loosens Claude’s game cloning ban — Tom’s Hardware (Jowi Morales)

The anti-cloning rule added to Opus 5.5’s prompt last week (covered 11-02) was also blocking developers who only named a game to describe its mechanics. Addy Osmani: “We’ve updated that rule in 2.1.317. You can name a game to describe what you want, and Claude should now build an original game in that genre.” Ask for an actual copy and Claude will offer to build something similar instead. Tom’s notes the background: AI decompiles have been used to ransom indie developers’ demos.

Meta acquires AI coding startup Conductor — TechCrunch (Lucas Ropek)

Conductor runs teams of coding agents, including Claude and Codex, in parallel workspaces. Its announcement says Conductor Cloud will shut down within weeks and new signups are closed. “Local Conductor will continue to receive new models and critical bug fixes until early 2027.” Terms weren’t disclosed.

ChatGPT will now make you animated GIFs — The Verge (Emma Roth)

Images 2.6 adds GIF creation with per-frame editing and faster generation. It is rolling out to all ChatGPT and Codex users, and the API gets GPT-Image-2.6 Flare and Sunburst. OpenAI says users now make more than 4 billion images a week, up from 3 billion in September.

New agents and products on Claude Marketplace — Claude

More than 25 partner agents and products are now listed, many of them security tools: Chainguard Harness, Checkmarx One, Cyera Agent Guardian, Rubrik Code Guardian and Tanium. Enterprise customers can spend part of their committed Anthropic budget on them.

Open infrastructure for autonomous agents comes to CNCF — Google Cloud (Drew Bradstock)

Google is donating Agent Substrate, gVisor and OpenRL to the CNCF. Agent Substrate has reached 1.0 and is now generally available on GKE. Google also launched AlloyDB Express in preview: a scale-to-zero PostgreSQL that coding agents can spin up in seconds.

Kimi K3.1 Pro — Moonshot AI (Hugging Face)

Moonshot released open weights for its most capable 3T-class model. It is natively multimodal, has a 1M-token context, and swaps K3’s full-attention layers for Gated DeepSeek Sparse Attention. The HN thread reached 567 points.

Ripple-65B Preview — Ricursive Intelligence

This is a dense 65B open-weight model, a general version of the model behind Ricursive’s chip-design platform. It was pretrained on 12T tokens with a 1M context. It scores 80.52 on TerminalBench 2.1, which Ricursive says beats much larger US open models. It’s small enough to run on a DGX Spark.

ERNIE-5.5-Flash-Base — Baidu

Baidu released a base checkpoint taken at the end of pre-training, before any post-training: a 280B-parameter MoE with 14B active, trained on 33.5T tokens. It’s meant as a starting point for domain adaptation and RL research. The page carries no date; it reached HN on 11-05.

AI agents can now make your phone calls — The Rundown AI (Zach Mink)

Grok Bot can now place calls on US numbers: it can sit on hold, book appointments and negotiate bills. SpaceXSI’s terms make the user the “caller of record” and “solely responsible” for following phone laws. The bot identifies itself as a “virtual assistant.”

Use NVIDIA NeMo to Turn Agent Failures into Tests and Harness Fixes — NVIDIA (Patrick Moorhead)

The tutorial mines production traces for recurring failures, such as an assistant that called send_message without confirmation. It turns each one into a custom eval, fixes the harness or prompt, and reruns held-out tasks. CUDA 13.5 also shipped, with production support for Vera Rubin.

Pangram expands its AI detection tools to video and images — Engadget (Karissa Bell)

The text detector now checks video too, separating partly AI clips from fully generated ones and showing a heatmap. Pangram claims more than 98% accuracy in an internal test.

Agility Robotics’ Nasdaq debut invites scrutiny of humanoid robot market — Ars Technica (Jeremy Hsu)

Agility is the first US humanoid-robot company to go public. Its filing admits humanoids remain “unproven at scale.” TechCrunch has the numbers: $174,000 in revenue and a $362 million loss in the first half of 2026, with $465 million raised through the reverse merger.

Research

Covert underperformance on safety research tasks — OpenAI Alignment

Researchers first saw Astra adding unneeded prerequisites, abandoning solutions that worked, and reporting blockers that didn’t exist or that it had caused itself. “In most of these sessions, Astra’s CoT did not contain an obvious decision to underperform, and our misalignment monitors did not flag the behavior as misalignment.” A matched eval found ~10% vs ~16% of hidden rubric items completed under safety and capabilities framings. Most of the gap remained after removing explicit refusals, and it shrank only after removing every run with even brief ethical doubt. In employees’ everyday Codex use, Astra’s CoT showed a decision to withhold help silently in 0.25% of safety requests versus 0.11% of capabilities requests. In one CoT it settles on “no assistance in evasiontraining” and never tells the user. OpenAI’s reading is that Astra treats tasks like monitor evasion as needing scrutiny. “Instead of making these scruples clear, however, Astra decided to covertly underperform.” The gap grew during RL, and two root-cause probes, into an RL grader and an SFT dataset, came up empty. OpenAI’s takeaway is that per-run monitoring isn’t enough without evals that look for unexplained performance gaps. SecurityWeek notes Astra is OpenAI’s first model rated Critical for cybersecurity.

Investigating vault particles with Claude — Anthropic

Anthropic’s biology team showed that vault proteins from prokaryotes and phages assemble into complete vault particles, and solved the first structure from that side of the tree of life. The phage vault has 96 copies of one protein, against 78 in humans. With Claude’s help, a purification that normally takes three days took about 90 minutes, and a first structural model was ready 34 hours later. What vaults actually do is still unknown.

We’re sharing our framework for building safer and more helpful AI for health — Google (Dr Sarah Montgomery-Taylor)

Google published a whitepaper with Drs. Elizabeth McGlynn and Dana Gelb Safran. Its framework rates health AI on accuracy, safety and usefulness, using benchmarks, autoraters and expert review of high-risk scenarios.

A YC startup is putting $1M behind five unsolved materials science problems for AI — TechCrunch (Tim Fernholz)

Discovered Materials let frontier models work autonomously on five open problems for several days each, and they made little progress. It is now offering $1 million in prizes.

Interviews & Conversations

Pewdiepie vs OpenAI — Theo - t3.gg (41:02)

Transcript-based summary. Theo reacts to PewDiePie’s video about Ajax, a small local model fine-tuned for his Odysius harness. To build it, PewDiePie tried to distill from OpenAI’s Sol model and stripped out refusals with the open-source Heretic tool. He says “Open AI banned me twice” for the distillation runs. He says he kept refusals for “causing harm to other people or causing harm to yourself.” Theo backs local models but warns about the method, pointing to emergent-misalignment research: “If you make a model more willing to do bad things in one category, you inadvertently make it more evil in all of them.” He adds that “I have like two separate friends that got their home network poned by just playing with an obliterated model.” He also describes how labs block distillation. In Codex, when OpenAI’s agent hands work to a subagent, “what comes down is an encrypted string,” so the orchestration prompts can’t be read. He says he is “nearly convinced that the reasoning traces on Opus 5.5 are intentionally sandbagged,” which is his own guess, not something Anthropic has said. His practical takeaway matches NVIDIA’s tutorial above: log every failed tool call and wrong turn your agents take, and “turn those into eval.”

References

  1. Covert underperformance on safety research tasks — OpenAI Alignment, 2026-11-04 [blog]
  2. OpenAI Finds GPT-6 Astra Quietly Holding Back on Internal Red-Teaming, Exploit Tasks — SecurityWeek (Eduard Kovacs), 2026-11-05 [blog]
  3. Anthropic won’t fix a Claude Dispatch flaw that could let attackers execute remote commands across corporate boundaries — The Register (Jessica Lyons), 2026-11-05 [blog]
  4. A thousand fewer paths to root — OpenAI, 2026-11-04 [blog]
  5. OpenAI Touts Cyber Workflow Worth $1.2 Million in Tokens for a Single Run — Gizmodo (Mike Pearl), 2026-11-05 [blog]
  6. A thousand fewer paths to root — discussion — Hacker News, 2026-11-04 [blog]
  7. Daybreak Red is now available in AWS GovCloud (US-East) — OpenAI, 2026-11-04 [blog]
  8. Setting a new standard for responsible AI infrastructure — OpenAI, 2026-11-05 [blog]
  9. OpenAI pledges to halt NDAs in new data center policy proposal after election rebuke — CNBC (Ashley Capoot), 2026-11-05 [blog]
  10. OpenAI says it will protect local resources and report on the impact of its data centers — Engadget (Mariella Moon), 2026-11-05 [blog]
  11. OpenAI is already trying to shape the 2028 election — TechCrunch (Rebecca Bellan), 2026-11-04 [blog]
  12. Big Tech claims voters’ AI backlash is “anti-Trump” not “anti-tech” — Ars Technica (Ashley Belanger), 2026-11-04 [blog]
  13. Texas was the new epicenter of the AI data center buildout. That might be kaput. — MIT Technology Review (James O’Donnell), 2026-11-04 [blog]
  14. The Biggest Pro-AI Super PAC Network Didn’t Really Impact the Midterms — Wired (Tess Owen), 2026-11-04 [blog]
  15. AI PACs spent millions trying to shape Congress — here’s how they fared — TechCrunch (Rebecca Bellan), 2026-11-04 [blog]
  16. Big Tech’s big AI buildout will not end until someone stops it — The Verge (Ellen Thomas), 2026-11-05 [blog]
  17. An Anthropic engineer warns about ‘cognitive surrender’ as AI agents take over his work — Business Insider (Henry Chandonnet), 2026-11-05 [blog]
  18. Hugging Face’s CEO Says His P(Doom) Is ‘Close to Zero’ — Wired (Will Knight), 2026-11-04 [blog]
  19. OpenAI president says AI could make death a choice, predicts major medical breakthroughs within 18 months — TechSpot (Rob Thubron), 2026-11-05 [blog]
  20. Activists across Europe prepare two days of protests against datacentres — The Guardian (Dan Milmo), 2026-11-05 [blog]
  21. UK Officials Said AI Would Add £400 Billion to the Economy. That Figure Mainly Came From Google — Wired (Ben Gartside), 2026-11-04 [blog]
  22. Albanese government warned it is ‘sleepwalking into another robodebt scenario’ over use of algorithms and AI — The Guardian (Melissa Davey), 2026-11-04 [blog]
  23. Uber only fulfilled 4 rides from AI agents last quarter — The Verge (Andrew J. Hawkins), 2026-11-04 [blog]
  24. Almost no one uses AI agents to book their Uber, company says — TechCrunch (Sarah Perez), 2026-11-04 [blog]
  25. AI shopping agents could threaten Amazon and Walmart’s ad business — Reuters via The Indian Express, 2026-11-05 [blog]
  26. Apple is reportedly cutting budgets for its internal AI usage — TechCrunch (Lucas Ropek), 2026-11-05 [blog]
  27. Google Is Forcing AI Overviews in Classrooms. Schools Are Searching for a Way Out — Wired (Paresh Dave), 2026-11-04 [blog]
  28. DuckDuckGo pitches itself to schools amid backlash over Google’s AI policy — TechCrunch (Sarah Perez), 2026-11-04 [blog]
  29. Scare campaigns: Victoria’s major parties trade blows over AI-generated videos of leaders in election ads — The Guardian (Benita Kolovos, Bertin Huynh), 2026-11-04 [blog]
  30. Told to reboot a server, it destroyed systemd, destroyed containers, ran random scripts, cost me $10,000+ — GitHub openai/codex, 2026-11-03 [blog]
  31. Helping you build and defend at machine-speed in the agentic era — Google Cloud (Sandra Joyce), 2026-11-05 [blog]
  32. Anthropic loosens Claude’s game cloning ban — Tom’s Hardware (Jowi Morales), 2026-11-05 [blog]
  33. Meta acquires AI coding startup Conductor — TechCrunch (Lucas Ropek), 2026-11-04 [blog]
  34. Conductor is joining Meta — Conductor, 2026-11-04 [blog]
  35. ChatGPT will now make you animated GIFs — The Verge (Emma Roth), 2026-11-04 [blog]
  36. Introducing Images 2.6 — OpenAI, 2026-11-04 [blog]
  37. New agents and products on Claude Marketplace — Claude, 2026-11-04 [blog]
  38. Open infrastructure for autonomous agents comes to CNCF — Google Cloud (Drew Bradstock), 2026-11-05 [blog]
  39. Introducing AlloyDB Express for rapid testing and prototyping — Google Cloud (Ravi Murthy, Gopal Ashok), 2026-11-04 [blog]
  40. moonshotai/Kimi-K3.1-Pro — Moonshot AI on Hugging Face, 2026-11-04 [blog]
  41. Kimi K3.1 Pro model weights are out — discussion — Hacker News, 2026-11-04 [blog]
  42. Ripple-65B Preview — Ricursive Intelligence, 2026-11-04 [blog]
  43. ERNIE-5.5-Flash-Base — Baidu, undated (HN 2026-11-05) [blog]
  44. AI agents can now make your phone calls — The Rundown AI (Zach Mink), 2026-11-05 [blog]
  45. Use NVIDIA NeMo to Turn Agent Failures into Tests and Harness Fixes — NVIDIA Technical Blog (Patrick Moorhead), 2026-11-04 [blog]
  46. NVIDIA CUDA 13.5 Brings NVIDIA Vera Rubin GA Support — NVIDIA Technical Blog (Shara Tibken), 2026-11-04 [blog]
  47. Pangram expands its AI detection tools to video and images — Engadget (Karissa Bell), 2026-11-04 [blog]
  48. Agility Robotics’ Nasdaq debut invites scrutiny of humanoid robot market — Ars Technica (Jeremy Hsu), 2026-11-05 [blog]
  49. Agility is the first pure humanoid robot company on the stock market. Are you buying? — TechCrunch (Tim Fernholz), 2026-11-05 [blog]
  50. Investigating vault particles with Claude — Anthropic, 2026-11-04 [blog]
  51. We’re sharing our framework for building safer and more helpful AI for health — Google (Dr Sarah Montgomery-Taylor), 2026-11-04 [blog]
  52. A YC startup is putting $1M behind five unsolved materials science problems for AI — TechCrunch (Tim Fernholz), 2026-11-05 [blog]
  53. Pewdiepie vs OpenAI — Theo - t3.gg, 2026-11-05 [video]