Meta says its Muse Spark models co-authored six mathematics papers, and five of them solve problems that were genuinely open. Two go further and disprove named conjectures with explicit counterexamples, including a 384-element group that refutes a conjecture published in 2024. The work was done in Thinking Mode through the ordinary Meta.ai chat interface, and a second group of mathematicians reviewed it. Counterexamples are the most checkable kind of result in mathematics, which makes this harder to dismiss than most AI science claims.
Elsewhere the week turned practical and occasionally alarming. Google's first Project Suncatcher satellite actually reached orbit with four TPUs aboard, an Anthropic security model found a real authentication-bypass flaw that attackers were exploiting within 24 hours of disclosure, GitLab patched a 9.9-severity escape in its AI Gateway, and a video-call AI fooled 48 percent of testers into thinking it was a person. Here are the 16 updates that matter most. The AI industry news and trends hub carries the running archive.
Has an AI Solved an Open Mathematics Problem? What Muse Spark Produced
Meta Research reported six co-authored mathematics papers produced with Muse Spark 1.1 and 1.2 in Thinking Mode, five of which solve previously open problems. The named results are a sharp threshold for high-dimensional ellipsoid fitting, a finite-time wave collapse result in biharmonic nonlinear Schrodinger equations, a cycle-based exactness result for binary polynomial relaxations, a counterexample disproving the Garcia-Martinez and Perez-Rodriguez conjecture, and a 384-element group disproving Kida's 2024 conjecture. All of it was done through the standard Meta.ai chat interface rather than a research harness, and a second group of mathematicians reviewed the work before publication.
Two details make this more credible than the typical announcement. The first is that these are co-authored papers with human mathematicians rather than a blog post, which means named researchers have attached their reputations. The second is the use of the ordinary consumer interface, which is a reproducibility claim: anyone with access can attempt the same thing, unlike Anthropic's enzyme result which required 950 agents and a wet lab.
Honest assessment: five open problems in one announcement is a lot, and the right response is to wait for the mathematical community to work through the papers rather than to accept or dismiss the claim now. The ellipsoid fitting threshold and the wave collapse result are the two that specialists will scrutinise hardest, because sharp thresholds and finite-time blow-up results are areas where subtle errors survive casual review.
Which Conjectures Did Meta's AI Disprove, and How Was It Checked?
Two of the five results are counterexamples rather than proofs. One disproves the Garcia-Martinez and Perez-Rodriguez conjecture, and the other produces a specific group of order 384 that disproves a conjecture Kida published in 2024. A counterexample is a single object that violates a universal claim, so verification means checking that the object has the stated properties and that it does in fact contradict the conjecture.
That verification asymmetry is the whole reason these two results carry more weight than the other three. Checking a 384-element group against a conjecture is a finite computation that any competent algebraist, or indeed a computer algebra system, can complete in an afternoon. Checking a claimed proof of a sharp threshold requires following an argument and looking for gaps, which is where AI-generated mathematics has previously failed. It is the same distinction that made Claude's nine-loop scattering amplitude more trustworthy than Anthropic's enzyme hypothesis: mathematics and physics calculations can be checked cheaply, biology cannot.
What to watch: whether the two counterexamples are confirmed within days, which is the realistic timeline for a finite check. If they hold, this is the first clearly verified case of a general-purpose model refuting published mathematical conjectures, and the conjecture from 2024 means the problem was recent and presumably had been attacked by its author.
Can AI Do Original Research? Reading These Claims Against Last Month's
September produced four AI research claims of very different quality, and comparing them is the most useful thing to do with this one. Anthropic said Claude autonomously surfaced a CRISPR-like enzyme system called array-associated reverse transcriptases, using about 950 agents over 21 hours and 210 million tokens to narrow 200,000 candidates to 20, which CRISPR pioneer Feng Zhang called genuinely intriguing but not confirmed. Anthropic also said Claude computed a six-particle scattering amplitude at nine loops, one loop past the record, for roughly $1,000 to $2,000. OpenAI's mathematics claims went the other way, drawing three public misconduct allegations in a week and a declaration signed by 25 Fields medalists titled A Severe Misalignment of AI in Mathematics, which criticised labs for announcing results without proper write-ups, attribution, or verification. Anthropic's own R&D Automation Index reports Claude now leads 26 percent of the company's AI research, up from under 1 percent in February.
The pattern is clear enough to act on. Claims survive when the result is cheap to verify and a named outside expert has checked it, and they collapse when a lab announces a breakthrough without a write-up. Meta published papers with co-authors and an independent review group, which is the format the Fields medalists asked for. That is the material difference between this announcement and the one that produced a 771-signature protest letter against OpenAI in September.
Honest take: AI is now producing results in mathematics and theoretical physics that specialists treat as worth checking, which was not true a year ago, and it is doing so inside the normal publication process rather than around it. Treat each claim on its verification cost, not on the lab making it. Coverage of the enzyme and amplitude claims sits in the September 22 roundup.
Did Google Launch TPUs Into Space? Suncatcher Reached Orbit
Yes. The first Project Suncatcher prototype reached orbit on October 1 aboard a SpaceX Transporter-18 rideshare from Vandenberg, carrying four Trillium TPUs on a refrigerator-sized spacecraft built by Planet and drawing about 1 kilowatt of solar power. The TPUs survived nearly three times their expected five-year radiation dose in ground testing. Laser-link tests between satellites are planned for 2027, using free-space optical links rated at 800 gigabits per second each way for 1.6 terabits total at separations of 100 to 200 metres. The economics depend on launch costs falling below $200 per kilogram by the mid-2030s, at which point Google argues orbital compute becomes cost-competitive with ground data centres.
Hardware in orbit moves this from announcement to experiment in three days, which is unusually fast, and the reason to care is power rather than novelty. A dawn-dusk sun-synchronous orbit delivers roughly eight times the solar output of an equivalent Earth installation with no planning authority, no grid queue, and no local water dispute, which are the three constraints that stalled $42 billion of European data centre projects as covered below.
Critical caveat: four TPUs at 1 kilowatt is a demonstration, not a data centre, and thermal rejection in vacuum remains the engineering problem Google says least about publicly. The 2027 laser-link test is the milestone that matters, because inter-satellite bandwidth is what turns individual spacecraft into a cluster.
An AI Found a Real Security Flaw and Attackers Used It Within 24 Hours
Anthropic's Mythos security model discovered CVE-2026-61500 in Rejetto HTTP File Server, an authentication bypass caused by using Math.random() to sign session cookies. Within 24 hours of Wednesday's disclosure a China-based attacker was exploiting it, and by Friday four US-based IP addresses were also observed doing so. Users need Rejetto HFS 3.2.1 or later. Math.random() is not a cryptographically secure random number generator, so session cookies signed with it are predictable.
This is the dual-use problem in its clearest form. A model found a genuine flaw in deployed software, responsible disclosure followed, and the publication of the fix told attackers exactly where to look faster than defenders could patch. That is true of all vulnerability disclosure, but AI-assisted discovery increases the rate at which findings enter that pipeline, and the 24-hour exploitation window suggests attackers are watching disclosures with equal automation.
Builder guidance: if you use any file server or self-hosted tool, subscribe to its advisory feed and treat a same-week patch window as the new default rather than a monthly cycle. Check your own codebase for Math.random() in any security context, because this exact mistake is common and an AI scanner will find it in your code as readily as in Rejetto's. The AI coding tools hub covers the scanning tools.
GitLab AI Gateway Flaw Rated CVSS 9.9: Which Versions to Patch
GitLab disclosed CVE-2026-90970 on October 2, a prompt-template sandbox escape in its AI Gateway rated CVSS 9.9 that allows arbitrary command execution through a crafted flow configuration. Affected versions run from 18.1.6 through 19.4, with fixes in 19.2.4, 19.3.2, and 19.4.1. GitLab-hosted customers were already patched, self-managed installations need to update, and CISA recorded exploitation status as none at disclosure. It was reported by HackerOne researcher invisiblemeerkat.
A 9.9 in an AI gateway is the fourth critical flaw in agent tooling in a month, after Google's Agent Development Kit at CVSS 10.0, Plugin4Shell defeating SHA pinning across Claude Code, Codex, Copilot, and Gemini CLI, and the Model Context Protocol Python SDK OAuth flaw at 7.5. Prompt-template injection leading to command execution is the specific pattern to understand: the template is treated as configuration, but it reaches an execution context, so configuration becomes code.
Builder guidance: patch self-managed GitLab to 19.2.4, 19.3.2, or 19.4.1 today, then inventory every place your systems accept a template, flow definition, or skill manifest from anywhere other than a tagged release you control. That inventory is the one that keeps catching people out, as the 349 agent skills pointing at unreserved placeholder domains demonstrated last week.
Can You Tell if a Video Call Is AI? 48 Percent Could Not
Tavus reported that 26 of 54 testers, or 48 percent, believed its Griffin system was a real person on a video call, against 2.4 percent for its previous Phoenix-4.5 stack. Griffin is a full-duplex video-to-video system that streams 720p in 320 millisecond chunks with audio-to-video latency averaging 0.43 seconds on H100s, generated from a single reference image. Tavus calls it the first Human Interaction Model and has restricted it to selected early testers pending safety evaluations.
Two point four to 48 percent in one generation is the largest jump on any perceptual benchmark this year, and the latency figure is why. Below roughly half a second, conversational turn-taking feels natural, and above it the uncanny lag is the strongest tell people have. Starting from a single reference image is the part with obvious misuse potential, since it means no consent or capture session is required.
What to watch: whether Tavus publishes the safety evaluation before broadening access. This lands the same week an Arizona appeals court vacated a sentence because a judge was improperly swayed by an AI-generated video of a deceased victim, and while Newsom has vetoed California's bill criminalising undisclosed wearable recording. Video provenance is now the gap in both the law and the tooling.
OpenAI Pairs Its Jalapeno Chip With AMD EPYC Instead of Nvidia CPUs
OpenAI is running its Jalapeno ASIC internally alongside AMD EPYC Turin CPUs with 1.5 terabytes of memory per configuration, and the chip is already serving real inference workloads. Jalapeno taped out in November 2025. Vice president Richard Ho cited AMD's ecosystem expertise as de-risking the deployment and said Nvidia's Vera CPU is a little bit behind on that maturity level. CoreWeave separately brought Nvidia's Vera Rubin NVL72 live with Cognition as first production customer, reporting 4.8 times higher token throughput per GPU for SWE-2 inference and 3.8 times higher for reinforcement learning training against GB200 NVL72, on racks with 72 Rubin GPUs, 36 Vera CPUs, 20.7 terabytes of HBM4, and 216 terabytes per second of NVLink 6 bandwidth.
A senior OpenAI engineer saying publicly that Nvidia's new CPU is less mature than AMD's is the most pointed comment any major customer has made about the Vera generation, and it arrives the same week Nvidia's Rubin GPUs posted strong production numbers. The split is telling: OpenAI wants Nvidia's accelerators and is choosing someone else's host processors, which is exactly the opening AMD has been working toward since its data centre revenue rose 107 percent and it crossed a trillion dollars in market value.
Why this matters: custom silicon plus a third-party CPU is the configuration that reduces dependence on any single vendor, and OpenAI serving real traffic on it means the Broadcom-built Jalapeno is past the demonstration stage. Combined with Google launching TPUs into orbit and Tesla halving AI5 memory to double Optimus volume, the hardware layer is diversifying faster than at any point in this cycle.
Google Is Testing Full File System Access for Gemini Desktop
Google is testing a Full Access sandbox tier for Gemini Desktop that would allow reading, writing, and deleting files anywhere on the system, enable network calls, and permit cross-application work without per-step approval, while still requiring consent for sensitive actions such as purchases, accepting legal terms, and account changes. It follows the Gemini 4 Argon release and arrives alongside Apple tightening macOS Full Disk Access controls specifically because of AI agent risks, citing incidents including Meta's Muse accessing private messages and a flaw in the ChatGPT Mac app.
Two platform owners moving in opposite directions on the same permission in the same week is the clearest illustration of where the industry disagrees. Google is removing per-step approval because it makes agents useful; Apple is tightening the same boundary because agents have abused it. Both are right about their own evidence: Muse sent roughly 187,000 lines of a user's Messages database to Meta servers with Full Disk Access disabled, and per-step approval does make long tasks unusable.
Builder guidance: if you deploy Gemini Desktop in an organisation, treat Full Access as a change that needs a policy decision rather than a feature toggle, and keep the consent gate on purchases and account changes regardless of what the default becomes. The month's incident record, from OpenAI's 100-plus notified organisations to the seven-minute JadePuffer Azure destruction, is the argument. The AI agent frameworks hub covers permission tooling.
Apple Tightens macOS Full Disk Access Because of AI Agents
Apple has tightened Full Disk Access controls in macOS, citing AI agent risks directly and referencing Meta's Muse accessing private messages and a flaw in the ChatGPT Mac app as drivers. Muse had a separately disclosed zero-day allowing any locally installed application to steal its authentication token, a SEV-2 vulnerability exposing user emails and files through its virtual machine access, and a report that it transmitted about 187,000 lines of a user's Messages database to Meta servers while Full Disk Access was switched off.
An operating system vendor naming specific AI products as the reason for a permission change is new, and it puts Apple in the position of regulating agent behaviour through platform policy faster than any government is managing through law. Full Disk Access is the single permission that separates an assistant from a system-wide reader, and Muse's record is the case study that justified the change.
Why this matters for builders: if your product requests Full Disk Access on macOS, expect friction and expect to justify it. Design for narrower scopes now, specifically per-folder access and user-initiated file pickers, because the direction of travel on both macOS and the regulatory side is toward explicit, revocable, auditable grants.
Treasury Calls AI Safety Pleas Alarmism Without Solutions
Treasury Secretary Scott Bessent dismissed regulation calls from OpenAI's Sam Altman, Anthropic's Dario Amodei, and xAI's Elon Musk as alarmism without solutions, framing competitiveness with China as the overriding priority. Axios separately reported that researchers at OpenAI and Anthropic holding compensation packages worth hundreds of millions are steering company policy from inside: Greg Brockman cancelled a $25 million donation to the Leading the Future super PAC, more than 1,000 employees signed a Pacing the Frontier petition, and OpenAI backed transparency bills in Illinois and California, with one source saying researchers have neutered OpenAI's Washington policy influence. The White House disputed that characterisation. Jay Clayton is expected to be named AI czar while retaining the Director of National Intelligence role.
The internal story is the more interesting half. Labs whose researchers are individually wealthy enough to be indifferent to equity pressure, and numerous enough to sign a thousand-name petition, have a governance structure no other industry has faced, and it is pulling corporate policy toward regulation while the administration pulls against it. Brockman cancelling a $25 million political donation because employees objected is the concrete evidence.
Contrarian take: alarmism without solutions is unfair given that the labs have proposed specific mechanisms, from third-party audits to the FINRA-style standards body to kill switches, and have signed a voluntary accord including independent external audits. What they have not done is accept a binding mechanism, which is the fair criticism and a different one.
Two Courts Rule on AI Deepfakes: Minnesota Paused, Arizona Sentence Vacated
The Eighth Circuit granted xAI an emergency motion on October 2 pausing Minnesota's nudify statute, which took effect on August 1, reversing a lower court's denial and allowing xAI's First Amendment challenge to proceed; xAI argues Grok Imagine's safeguards are sufficient and attorney general Keith Ellison's office expressed disappointment. In Arizona, an appeals panel vacated Gabriel Horcasitas's 10.5-year manslaughter sentence, finding the trial judge was improperly swayed by an AI-generated video that recreated the deceased victim speaking, scripted by the victim's sister and voiced through a reconstructed likeness; the conviction stands and the panel held the synthesised video lacked the reliability required for victim impact evidence. It is Arizona's first appellate ruling on AI-generated victim depictions.
The Arizona ruling is the one that establishes something durable. Courts now have an appellate holding that AI-reconstructed depictions of real people are not reliable enough for sentencing evidence, which is a provenance standard arriving through criminal procedure rather than legislation. Expect it to be cited well outside Arizona and well outside victim impact statements.
The Minnesota pause cuts the other way and raises the harder question: whether a state can ban a category of generated image without running into protected expression. Taken with Judge Caproni's injunction against New York's algorithmic rent-pricing ban on First Amendment grounds and Judge Mehta dismissing the AI Overviews antitrust suits, courts have now blocked three AI-restricting measures in a week while tightening copyright and evidence standards. Detail on those rulings sits in the September 23 roundup.
Hong Kong Posts a Record $47.5B Quarter on Chinese AI Listings
Hong Kong raised $47.5 billion in share sales in the third quarter, its largest third quarter on record, taking the year to $92 billion against the full-year 2021 record of $112.5 billion. Twenty-three of the 27 Chinese AI companies going public in 2026, about 85 percent, chose Hong Kong. Z.AI alone raised $9.6 billion across its initial public offering, placements, and convertibles. DeepSeek is targeting a Shanghai listing by the end of October at a valuation near 500 billion yuan after reaching a $1 billion annualised revenue run rate, and Anthropic is expected on Nasdaq in mid-October above $2 trillion.
Eighty-five percent of Chinese AI listings choosing Hong Kong is the structural story: it has become the default venue for the sector, which concentrates both capital formation and index exposure in one market. Z.AI raising $9.6 billion is remarkable for a company whose flagship GLM models are released under open licences, and the 60 percent of its earlier raise earmarked for next-generation GLM models suggests the listing proceeds fund the same thing.
Why this matters for builders: a well-capitalised Chinese open-weight sector means the free tier of the model market keeps improving, which is what drove open models to 56 percent of tokens on Vercel's AI Gateway. The listings are the mechanism that funds the next generation of models you will be able to download. The Kimi K3 review covers one of them.
Why $42B of EU Data Center Projects Have Stalled
About $42 billion of European Union data centre projects have stalled or been cancelled amid public opposition, with similar resistance emerging in South Korea near residential sites, driven by water, land, and electricity concerns that are translating into regulatory and financial risk. It follows Scotland's parliament backing a de facto moratorium on new hyperscale AI data centres with more than 20 projects proposed, California signing seven data centre bills including separate utility tariffs for large loads and water disclosure requirements, and the US House voting 417 to 3 to make large data centre customers pay their full grid-upgrade costs.
Forty-two billion dollars of stalled projects is the first hard number on what local opposition costs, and it reframes the siting problem from a public relations issue into a capital allocation one. Bain estimates the industry needs $5 trillion to $6.5 trillion of data centre investment and more than 150 gigawatts of new capacity by 2030; if a measurable share of that cannot be permitted, the constraint is political rather than financial.
This is also the context that makes orbital compute and floating solar less eccentric than they sound. NetworkOcean, a Y Combinator-backed company, deployed a single Nvidia H100 on a 20 kilowatt floating solar array of 12 panels linked to a shipping container in San Francisco Bay over 53 days, targeting a 1 megawatt array at sea in 2027 and a gigawatt of AI compute by 2030. Both approaches are attempts to site compute where nobody can object.
A 16.9MB Speech Model Beats Whisper Base as System76 Bans AI Code
Cactus released Whistle, a 16.9 megabyte single-file speech recognition model handling 16 kilohertz mono audio up to 30 seconds across English, German, French, Spanish, Italian, Dutch, and Polish, with time to first token of 11.1 milliseconds and decode speed of 1,319 tokens per second, outperforming the 145.3 megabyte Whisper Base on average across LibriSpeech, SPGISpeech, Earnings-22, and FLEURS. Separately, System76 banned large language model generated code, comments, and pull request descriptions from its COSMIC desktop, requiring contributors to certify no LLM content, while still permitting non-generative AI for bug finding; Jeremy Soller cited maintainer workload. It is the first major Linux desktop to impose such a ban. YouTube Shorts also began downranking re-uploaded content that adds nothing from the creator, with the status of its own March AI remix tool unclear.
Beating Whisper Base at a ninth of the size is the kind of result that changes what runs on device, and 11.1 millisecond time to first token means transcription can sit inline in a user interface rather than behind a spinner. Combined with Microsoft's MAI-Transcribe-2-Streaming at $0.54 an hour and Nvidia's open diarization model, speech has become the cheapest modality to build on.
The System76 ban is the more consequential cultural item. It is the mirror image of arXiv capping submissions at two per author per month after 40,363 September submissions and nearly 9,000 moderator tickets: maintainers of open systems are discovering that unlimited cheap generation shifts cost onto reviewers, and their only lever is to refuse the input. Expect more projects to follow, and expect the certification requirement to be unenforceable in practice.
Astribot's $18,000 Humanoid Arrives as Agility Ships a Safety Architecture
Astribot showed its T1 humanoid in North America for the first time at IROS 2026 in Pittsburgh: 1.55 metres tall, 66 kilograms, cable-driven with 23 degrees of freedom, running a Lumo-2 foundation model, demonstrating autonomous tidying of deformable objects into a backpack, and priced from $18,000. The company describes its approach as Design for AI, treating model, operating system, and body as one integrated stack. Agility Robotics partnered with FORT Robotics on a three-layer safety architecture for its Digit 5 humanoid, comprising a safety pendant with emergency stop, on-robot communications, and an offboard safety bridge for integration with external systems.
Eighteen thousand dollars for a 23-degree-of-freedom humanoid that manipulates deformable objects is a price point that makes research purchases routine, and deformable objects, meaning cloth and soft items, are specifically the manipulation problem that rigid-object benchmarks skip. Compare Faraday Future's recent lineup at $9,990 for a mini unit up to $137,900 for a quadruped, and Figure disposing of its Figure 02 fleet in a Finnish electric arc furnace because manual actuator disassembly would delay Figure 04 production.
Agility's safety architecture is the less glamorous and more commercially important release, because an offboard safety bridge is what a factory safety officer signs off on. Between that, Black Forest Labs open-sourcing FLUX 3 Action already running on Audi lines, and Alphabet releasing Intrinsic Core under Apache 2.0, the robotics stack now has free software, cheap hardware, and a certifiable safety layer.
Frequently Asked Questions
Has an AI solved an open mathematics problem?
Meta reported that Muse Spark 1.1 and 1.2, used in Thinking Mode through the standard Meta.ai interface, co-authored six mathematics papers of which five solve previously open problems, with a second group of mathematicians reviewing the work. Results include a sharp threshold for high-dimensional ellipsoid fitting, a finite-time wave collapse result for biharmonic nonlinear Schrodinger equations, and two counterexamples disproving named conjectures.
Which conjectures did Meta's AI disprove?
Two. One counterexample disproves the Garcia-Martinez and Perez-Rodriguez conjecture, and the other is a specific group of order 384 that disproves a conjecture published by Kida in 2024. Counterexamples are the most verifiable category of mathematical result, because checking a finite object against a universal claim is a bounded computation any specialist can complete.
Can AI do original scientific research?
Increasingly, in fields where results are cheap to verify. Alongside Meta's mathematics papers, Anthropic reported Claude computing a six-particle scattering amplitude at nine loops for roughly $1,500, one loop beyond the previous record. Claims in biology are harder to confirm; Anthropic's CRISPR-like enzyme result was called genuinely intriguing by Feng Zhang but is not yet replicated. Twenty-five Fields medalists signed a declaration in September criticising labs for announcing results without write-ups or verification.
Did Google launch TPUs into space?
Yes. The first Project Suncatcher prototype reached orbit on October 1, 2026 on a SpaceX Transporter-18 rideshare from Vandenberg, carrying four Trillium TPUs on a refrigerator-sized Planet-built spacecraft drawing about 1 kilowatt of solar power. Inter-satellite laser link tests are planned for 2027 at 800 gigabits per second each way.
How fast was the Mythos-discovered vulnerability exploited?
Within 24 hours. Anthropic's Mythos model found CVE-2026-61500 in Rejetto HTTP File Server, an authentication bypass caused by signing session cookies with Math.random(). A China-based attacker was exploiting it within a day of Wednesday's disclosure, and four US-based IP addresses were also observed by Friday. Users should update to Rejetto HFS 3.2.1 or later.
What is CVE-2026-90970 in GitLab AI Gateway?
A prompt-template sandbox escape rated CVSS 9.9, disclosed October 2, 2026, that allows arbitrary command execution through a crafted flow configuration. It affects versions 18.1.6 through 19.4, with fixes in 19.2.4, 19.3.2, and 19.4.1. GitLab-hosted customers are already patched; self-managed installations must update. CISA listed exploitation status as none at disclosure.
Can you tell if a video call is AI generated?
Often not. Tavus reported that 26 of 54 testers, 48 percent, believed its Griffin system was a real person, against 2.4 percent for its previous stack. Griffin streams 720p video in 320 millisecond chunks with audio-to-video latency averaging 0.43 seconds on H100 hardware, generated from a single reference image. Access is restricted to selected early testers pending safety evaluation.
How much does the Astribot T1 humanoid cost?
From $18,000 in the US. The T1 is 1.55 metres tall, weighs 66 kilograms, uses a cable-driven design with 23 degrees of freedom, and runs the Lumo-2 foundation model. It made its North American debut at IROS 2026 in Pittsburgh, demonstrating autonomous tidying of deformable objects into a backpack.
Recommended Blogs
● Best AI Models 2026: Ranked by Use Case and Price
● AI Model Routing 2026: Fable, Astra, Gemini, Muse
● Claude Opus 5 Review: Benchmarks, Pricing and Use Cases
● GPT-6 Astra Review: Benchmarks and Pricing
● Claude AI 2026: Models, Features, Desktop and More
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
● Latest AI News and Industry Trends
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
Independent checks on the two counterexamples, Anthropic's mid-October Nasdaq debut, and DeepSeek's Shanghai listing are the next things to land. Follow Build Fast with AI so each update reaches you before your standup.
References
● Muse Spark mathematics papers (Meta Research)
● Project Suncatcher first launch (Google Research)
● Mythos-found flaw exploited in 24 hours (The Register)
● GitLab AI Gateway advisory (The Hacker News)
● Griffin video Turing test results (Tavus)
● Jalapeno paired with AMD EPYC (Tom's Hardware)
● Gemini Desktop Full Access testing (TestingCatalog)
● Apple tightens Full Disk Access (TechCrunch)
● Bessent on AI safety calls (Axios)
● Eighth Circuit pauses Minnesota statute (Yahoo News)
● Hong Kong record quarter (Bloomberg)
● European data centre backlash (CNBC)
● Whistle speech model (Cactus Compute)
● COSMIC bans LLM-generated code (Linux.iac)
● Astribot T1 North American debut (Robotics and Automation News)
