October 3rd 2026
Curated AI news and stories.
Amazon reportedly seeks outside investors for $8 billion of Nvidia chips
Amazon is exploring a transaction that would move about $8 billion of Nvidia chips into a separate financing entity while retaining use of the hardware, Bloomberg reported Friday, citing the Financial Times. The proposed would hold Grace chips across more than a dozen US data centers, issue debt and potentially sell investors an equity stake of up to 10%.
The proposal would change who finances the equipment without establishing a reduction in Amazon's computing demand. It remains a reported financing plan. The account reviewed does not establish a completed transaction or the guarantees investors would receive. SourcesB
A phishing campaign impersonated AI insiders to target US policy experts
Proofpoint disclosed Thursday that a group it tracks as TA419 impersonated an Anthropic employee and prominent policy figures to target experts at US think tanks, universities and law firms. The February and July campaigns used invitations to discuss military AI or contribute to policy work as the opening for theft.
The attackers relayed genuine Microsoft sign-ins through a fake interface, capturing authenticated sessions even when victims supplied a second authentication factor. Proofpoint assesses the group as China-aligned and motivated by espionage. That attribution is the security firm's assessment; the report does not establish that Anthropic participated or that its own systems were breached. SourcesA
Google moves private training into protected servers and exposes the access rules
Google Research announced Friday a system already adopted by Gboard that uploads encrypted training examples to protected server environments. Devices authorize which computations can access their data, and the access policies go into a public transparency log. The system releases anonymized model updates.
This changes the privacy question from whether training happens on a phone to whether outsiders can verify what server code may do with its data. Google says the approach improves training speed and accuracy. The guarantees depend on , hardware-isolated areas whose current limitations Google explicitly acknowledges. A protected server is still a security dependency. SourcesA
Inworld buys Ultravox to bring voice-agent conversations into its infrastructure
Inworld announced Thursday that it acquired Ultravox, the voice-agent platform formerly known as Fixie. Joining Ultravox staff will continue developing the platform. The purchase combines Ultravox's speech understanding and handling of conversational interruptions with Inworld's speech models and infrastructure.
For developers, the change puts the platform that manages a live conversation and the infrastructure that produces its voice under the same owner. GeekWire reported Friday that the purchase price was undisclosed. Neither the announcement nor the reporting establishes an independently measured improvement in conversation quality from the combination. SourcesAB
OpenAI's internal Jalapeño deployment pairs its custom chip with AMD hosts
OpenAI is deploying its Jalapeño chips internally alongside AMD EPYC Turin host processors, Tom's Hardware reported Friday. Hardware chief Richard Ho told the outlet the team chose the platform to reduce design risk and draw on its partners' experience. He described Nvidia's standalone Vera offering as less mature for that decision.
The report adds a concrete host-platform choice and internal deployment to OpenAI's earlier rollout plans. It does not identify a OpenAI service running production traffic on Jalapeño. Internal deployment alone cannot establish that customer-facing milestone. SourcesB
doxx.net opens a private-network beta for people and their agents
doxx.net announced Thursday an open beta and a $38 million led by Andreessen Horowitz. Its platform lets people and establish private networks, with controls over connectivity and blocking of malicious destinations through the internet's domain-name system.
The strategic claim is that an agent's network access can be constrained outside the model making its decisions. That makes the launch worth examining beyond the funding round. The company's threat-blocking and privacy claims have not been independently tested here, and blocking known malicious destinations does not establish protection against every unauthorized action an agent might take. SourcesA
A mathematician says half of Meta's six "open problem" papers were already solved
Meta published six papers on Friday in which mathematicians worked with Muse Spark 1.1 and 1.2 in Thinking Mode through the ordinary meta.ai chat interface, and said five of them answer previously open questions. They span probability, PDEs, group theory, optimization, arithmetic physics and non-associative algebra. One proves finite-time blow-up for a class of radial wave-equation solutions, a question left open in 2015. Another reports a 384-element counterexample to a 2024 conjecture, found with search code the model wrote for the GAP algebra system. Meta says a second group of mathematicians reviewed each paper and that the papers mark which passages the AI drafted.
On Saturday, mathematician Jason Dean Lee said three of the six had already been resolved by others. Meta's own post concedes part of this: it credits three concurrent August works on Gaussian ellipsoid fitting, a September 16th group-theory result from an AI agent called Nilradical, and independent work by Hu and Wen on evolution algebras. What is in dispute is whether "open" was the right word for those problems when the papers appeared. Meta had not answered Lee's count in anything found this afternoon, and the second-hand write-up of his post is the only account of it. Meta's claim for the remaining problems stands on its reviewers, not on an independent check. SourcesAC
Anthropic commits $100 million to train 10,000 engineers to deploy Claude inside companies
Anthropic announced the Claude Frontier Academy on Friday, with a goal of 10,000 "Frontier Deployed Engineers" by the end of 2027. Each trainee does a multi-day in-person program that ends in a graded practical on a simulated enterprise deployment, then a 12-week residency leading a real Claude project at their own employer with Anthropic engineers supporting. First cohorts run in San Francisco, New York and London and include engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. Graduates get a badge, with the first expected in early 2027.
Entry is by nomination through an Anthropic account team, and the announcement names no price for participants. That limits it to customers large enough to spare an engineer for twelve weeks. Anthropic's stated reason is that the scarce input in enterprise AI is people who can deploy it safely and the people it trains stay on their employers' . SourcesA
Salesforce agrees to buy Listen Labs, an AI customer-research startup
Salesforce signed a definitive agreement this week to acquire Listen Labs, whose agents design studies, recruit participants, run interviews in more than 120 languages and simulate buyer reactions with "digital twins" built from real customer behavior. Salesforce has not disclosed a price. Business Insider reported about $2 billion three weeks ago, which is unconfirmed. Listen Labs raised a $69 million Series B in January at a $500 million , and was reported to be near $30 million in annual revenue during sale talks, a figure nobody has confirmed. Customers include Microsoft and Anthropic.
The technology goes into Marketing Cloud and Service Cloud, and co-founder Alfred Wahlforss and CTO Florian Juengermann join Salesforce AI Labs. The deal is expected to close in the fourth quarter of Salesforce's fiscal 2027, pending regulatory approval. SourcesBB
DeepSeek's Harness gets a desktop app with plugins and scheduled jobs
DeepSeek released a v0.2 preview of Harness, its open-source agent , as a desktop app for macOS (Apple silicon and Intel) and Windows, with Linux through an package. Version 0.2 bundles its own Node runtime, adds an in-app plugin manager where users install extensions by typing an npm package name, schedules automation tasks on a timer, and ties web search to the working session. It is MIT licensed, but it needs a DeepSeek account or , and the supported model is DeepSeek-V4.1-Flash.
The scheduled tasks are the notable addition: a coding assistant becomes something that runs jobs while its owner is away, which is the permission problem every other agent desktop is currently dealing with. Outlets disagree on the date, with one placing the release on September 29th and DeepSeek's own account posting about it October 2nd, so treat the timing as unsettled. The beat otherwise produced no new Chinese model release this afternoon. SourcesBB
Gemini's desktop app has an unreleased setting that lets it act outside chosen folders
TestingCatalog found a new "Additional options" control in the section of Google's Gemini desktop app, which would let the agent work outside selected folders, reach the internet and operate across other apps without asking at each step. Purchases, account creation, legal terms and sensitive edits would still require approval. The setting was seen on September 30th and reported October 2nd.
Google has not confirmed the feature or given a release date. The capability appears under trusted testing, and nothing links it to Gemini 4 Argon, which launched the same day to a limited group of cyber defenders. This is a find in unreleased code, and builds change before shipping. SourcesC
Cactus ships a 16.9 MB speech-to-text model that runs on a CPU
Cactus Compute released Whistle on October 2nd, an open speech-to-text model in a single 16.9 MB file. It transcribes up to 30 seconds of 16 kHz mono audio per pass in English, German, French, Spanish, Italian, Dutch and Polish, and returns word-level timestamps with probabilities and speech . The company reports 11.1 milliseconds to first on an Apple M4 Pro and 1,319 tokens per second against Whisper Base's 266, with better accuracy than Whisper Base (145.3 MB) and Moonshine Tiny v2 (41.9 MB) on LibriSpeech, SPGISpeech and Earnings-22.
Those are the vendor's own measurements on public , and the 30-second window means longer audio has to be chunked. Weights are on and the code is on GitHub, installable with pip install cactus-needle. SourcesA
Trump is expected to name intelligence chief Jay Clayton as AI adviser
President Trump is expected to choose Director of National Intelligence Jay Clayton as his top AI adviser, Reuters reported Friday, citing a source familiar with the decision. Clayton is expected to retain his intelligence role. The White House said personnel announcements would come from the president and dismissed reporting before then as speculation.
The appointment is therefore still a reported plan. It would put the administration's AI coordination alongside intelligence responsibilities, following David Sacks's departure from the formal adviser role. Neither a confirmed appointment nor a detailed mandate was established in the reporting reviewed. SourcesB
AWS discloses agent-control takeover and credential exposure in Loom
AWS's Friday bulletin describes three flaws in Loom, its open-source agent orchestration platform. A deployment without an could give a network caller administrative authority, including access to integration credentials and permission policies. Other flaws let authorized integration editors redirect credential-bearing requests or reach internal network addresses.
AWS recommends version 1.7.0 and patching derivative code. Its instructions also include rotating integration secrets and revoking affected access tokens. The authentication flaw had already been fixed in version 1.6.1 in August; Friday's disclosure is not evidence that every fix was released Friday. SourcesA
SageMaker users need to restart Spaces to receive a security fix
AWS disclosed Friday that improperly sanitized connection details could let code run in another project member's SageMaker Space. Where is enabled, a contributor could gain access to that member's temporary credentials and use downstream services as that person.
AWS says the fix is deployed globally for supported versions and takes effect at the next Space startup. It recommends restarting affected Spaces. Some older version lines are out of support and have no fix. A shared project membership is therefore part of the exposure boundary; this is not a claim that any internet user could exploit the flaw. SourcesA
Meta opens the software that connects Muse to homemade hardware
Meta's Muse Gadgets repository supplies firmware and device SDKs for boards and Linux computers, including Raspberry Pi. Builders can connect displays, sensors and controls to Muse. Pairing requires an token and developer mode in the mobile app.
The repository uses Apache-2.0 with stated exceptions for third-party components and its avatar. That opens the device software for modification; it does not remove the Muse pairing requirement or the separate SDK terms. The published code was inspected, but no hardware was flashed or tested. SourcesA
The parallel run also caught the distribution move around the code: Meta is giving 5,000 Muse subscribers a free USB-C device called Muse Home Link for home networks and smart devices. The useful signal is not the board itself. It is Meta courting the Raspberry Pi and ESP32 crowd while Apple is tightening agent permissions on the Mac. SourcesBB
Cloudflare opens Qwen-derived models that return decisions instead of prose
Cloudflare has published Clef and Clef-Flash, with downloadable under Apache-2.0. Their model cards identify Qwen backbones of 27 billion and 9 billion parameters. Each takes a state and typed questions, then returns probabilities over permitted answers without generating an open-ended response.
The supplied interface is compatible with Jev and SystemOne. That gives builders a self-hosted alternative for classification and routing with a constrained output format. Cloudflare's comparison scores are internal measurements, and a probability field does not by itself establish that the confidence is calibrated on a customer's data. SourcesAA
The same interface is also landing in local runtimes. llama.cpp added a /v1/systemone for decision models, and Ollama shipped the same in v0.35.0. That turns routing and triage into millisecond-scale local calls on models ranging from Julia-1 at 144 million to the multimodal OpenJev at 27 billion. SourcesAA
Nvidia prices a 64GB DGX Spark at $4,999
Nvidia announced Friday that a 64GB DGX Spark will ship through hardware partners on October 23rd, starting at $4,999. The configuration keeps the GB10 platform and Nvidia software stack. The company says two units can connect through their built-in networking to pool 128GB of memory.
Nvidia advertises support for models up to 100 billion parameters on one unit. That ceiling is a vendor claim, not a guarantee about a particular model's , context length or speed. Buyers deciding whether to host an agent locally still need measurements for the workload they intend to run. SourcesA
The parallel run preserved the price comparison the official post omits: coverage from hardware outlets says the 128GB original launched around $4,000 and now lists at $6,950, while the new 64GB version starts at $4,999. Read that as memory rationing. The AI desktop is being segmented around RAM before it is being segmented around compute. SourcesBB
A Tokyo ruling recognizes protection for an actor's cloned voice
Tokyo District Court recognized legal protection for Kenjiro Tsuda's voice in a ruling Wednesday, AP reported Thursday. Tsuda had sought removal of TikTok videos using an unauthorized AI imitation of his voice. The development changes the dispute from an awaited judgment into a court decision.
AP describes it as Japan's first acknowledgment of a right to one's voice. The report does not establish a general ban on voice synthesis, and the judgment itself was not available for review. SourcesB
AWS's security-scanning MCP server could write outside its workspace
An October 1st AWS bulletin identifies an argument-injection flaw in security-agent-mcp-server's differential scanning operation. A crafted reference could be interpreted as a command-line option, allowing files outside the intended workspace to be created, overwritten or truncated.
AWS says versions from 0.1.1 through those before 0.2.0 are affected and recommends upgrading. The practical failure is in the tool's enforcement of its filesystem boundary. A model choosing the correct scanning tool would not make that boundary safe. SourcesA
Alibaba Cloud schedules a DeepSign model identifier for retirement
Alibaba Cloud's September 23rd notice says AI DeepSign will stop supporting the deepseek-v4-flash identifier on October 9th, Beijing time. It names deepseek-v4-flash-0731 as the replacement and warns that tasks still using the retired identifier will fail to return results.
This is a service-specific migration deadline. It does not establish the retirement of DeepSeek's own API or the withdrawal of downloadable weights. Purchased DeepSign packages and credit balances are unaffected, according to Alibaba. SourcesA
ScholarCatalyst finds research agents miss papers that inspired real projects
An October 1st builds a retrieval benchmark from lead authors' judgments about which earlier papers helped, or could have helped, their research. The task restricts the searchable literature to what existed when each project began.
The authors report that agentic search retrieved fewer relevant papers than embedding search in their comparison. Even a stronger agent recovered only about half the target papers among its first 20 results. The test exposes a gap between finding text related to a question and finding the prior idea that makes the question tractable. Retrospective author judgments remain a limitation. SourcesA
VISTA preserves visual observations instead of compressing them into notes
The VISTA technical report, submitted October 1st after an August blog version, gives a multimodal agent a retrievable visual memory that preserves previous observations. The model can reorganize those observations while working through interactive tasks.
The researchers report a perfect Relative Human Action Efficiency score on the 25 public games with Claude Opus 5.0. This is a public-game result from the authors, not a private-test score or proof of general intelligence. The useful finding is that changing the surrounding memory system changes what the same model can complete. SourcesA
A memory study shows why never-retrieved facts can be mistaken for useless ones
Causal Memory Policy, an October 1st preprint, identifies a blind spot in agents that judge memories by their effect on completed tasks. A memory that never gets retrieved cannot demonstrate its value. Deleting or retaining it then produces the same observed result.
The researchers reserve context slots for sampled memories to measure that missing effect. Their experiments improve discrimination between useful and unnecessary memories, but the value measured for one question does not reliably predict usefulness on unseen questions. The result challenges automatic deletion rules that treat lack of observed benefit as evidence of irrelevance. SourcesA
Mingbird reports better task completion from small models with stricter controls
Mingbird's October 1st preprint presents an agent framework for Windows and Ollama that constrains context growth, detects repeated and checks the original task before accepting completion. The authors report higher completion scores than the compared frameworks while holding models, machine and budgets fixed.
They also disclose limits: a benchmark they built themselves, one machine and single-trial scoring. Repeated runs move enough to weaken claims about individual components. The framework is worth testing for local deployments, but its published result is not a universal ranking of agent products. SourcesA
Irrigation agents improve when software retains control of physical constraints
Mimir, described in an October 1st preprint, makes an propose irrigation actions while a simulator checks and revises them before execution. Lessons from recurring failures can change future proposals, but the physical model and execution constraints remain fixed.
In retrospective evaluation across sites, crops and years, the authors report about 51% less irrigation than historical schedules. That is not evidence from a prospective farm deployment. The architecture offers a concrete division of responsibility: the model can revise its advice without acquiring authority to rewrite the rules that constrain the equipment. SourcesA
Robot replanning does not reliably undo an action error
An October 1st study injects small action errors into robot manipulation tasks and measures whether they shrink or grow. It compares replaying a sequence of actions with replanning after the disturbance.
Across 12 tasks, the authors find that replanning rarely turns error amplification into confident contraction. Their results question the assumption that frequent new decisions automatically make an imitation-trained robot recover. They argue for training that deliberately exposes policies to perturbations and the recovery actions those perturbations require. These are benchmark findings, not incident rates for deployed robots. SourcesA
GUI agents learn different lessons from training and from temporary context
An October 1st preprint splits GUI-agent experience into components rather than feeding whole interaction histories into either training or memory. In its experiments, recurring locators and lessons worked better when learned into model weights, while procedures and current-state facts worked better in context.
The authors test the separation across model families and environments, reporting improvements over sending everything to one destination. The deployment question is whether an experience describes a durable skill or a situation that will soon change. Code and data are promised, so independent reproduction remains outstanding. SourcesA
A smaller observer can identify where another model starts hallucinating
Researchers report an approach to locating the beginning and continuation of hallucinated passages by examining models' internal representations. In their October 1st preprint, an external observer sometimes matches or exceeds a generator's ability to detect its own errors, even when the observer is smaller.
The result suggests that self-assessment need not set the ceiling for automated checking. It does not establish a general fact-checker: the experiments detect marked spans under a particular evaluation setup, and inspecting internal states requires access that many hosted services do not provide. SourcesA
A privacy result puts conditions on verifying AI work without revealing its inputs
An October 1st theoretical paper asks whether AI outputs can be verified without exposing confidential information when their correctness depends on outside answers, such as human judgments or experiments. It proves that a general zero-knowledge guarantee is impossible in the model it studies.
The positive result requires those outside answers to carry cryptographic signatures, alongside the paper's . This is not a finding that all private AI auditing is impossible. It identifies a missing condition that designers cannot replace with a generic promise to verify without seeing the data. SourcesA
JevSpawn adapts the menu of actions a decision model can choose
JevSpawn, an October 1st preprint, connects natural-language tasks to finite sets of possible actions. It generates alternative branches, selects among them using feedback, revises the representation and retains alternatives for recovery.
The researchers report improved performance and faster navigation in their benchmark comparisons. The work addresses a concrete limit of decision models: choosing quickly from a supplied menu is useful only if the menu contains the next action the task needs. The published evaluation has not been independently reproduced here. SourcesA
A long-task agent tracks unfinished requirements alongside its view of the world
An October 1st preprint introduces PoS, a framework that maintains explicit beliefs about current conditions and unresolved task requirements. It checks those beliefs for consistency and detects cases where an agent keeps acting without making progress.
The authors report gains across four execution and diagnosis benchmarks with three model backbones. Their proposal makes completion depend on maintained requirements and recovery checks, rather than merely retaining more conversation history. Whether those checks hold up as tools and environments change remains a deployment test. SourcesA
PyPottery applies image models to the archaeology publication backlog
PyPottery's October 1st paper describes an open-source suite for extracting pottery drawings, reading handwritten annotations, inking pencil sketches, vectorizing them and preparing publication layouts. The evaluation uses 240 drawings on 50 sheets from an Italian archaeological site.
Participants reported a median perceived speedup of 40 times. That is a usability-study response, not an independently timed measurement. The project nevertheless offers a specific use for image models in documentation work that often delays publication long after an excavation ends. SourcesA
OpenAI's safety-systems lead resigns days after three researchers were fired
David Robinson, who led OpenAI's Safety Systems team and worked on the system cards that document its models' tested behavior, has resigned, Business Insider reported Friday. His exit lands days after the company dismissed three safety researchers for what a spokesperson called mishandling sensitive information outside established procedures.
Subsequent reporting named the three as Jasmine Wang, Tomek Korbak and Mikita Balesni, all of whom worked on and safety. Korbak wrote in late September that he was OpenAI's technical contact for 's investigation of the Hugging Face intrusion, and the Wall Street Journal reported the information at issue went to an outside safety organization. OpenAI is simultaneously answering a California subpoena over its agents' incidents. The company keeps saying it has internal channels for safety concerns; this week it fired three people and lost the person who documented the models. SourcesBB
Aleph Alpha releases Kolibri-1, an Apache-licensed model built to reason in German
Aleph Alpha published Kolibri-1 on Hugging Face Saturday under an license: a 78-billion-parameter model with 3.46 billion active per token, 384 experts per layer with six routed, a of 1,048,576 tokens and selectable reasoning effort. The company pitches it as a sovereign model for German government and industrial buyers, built to reason in German rather than translate through English, and to run on infrastructure customers control.
The release comes months after Cohere agreed to acquire Aleph Alpha in a transatlantic combination. An Apache-licensed frontier-adjacent model aimed at European public-sector deployment is a different bet than another API: the license is the product. A procurement officer still has to establish whether a model with 3.46 billion holds up against Chinese open-weights leaders running ten times that. SourcesAB
Employers add 29,000 jobs in September, and information sheds another 10,000
The Bureau of Labor Statistics reported Friday that US payrolls grew by 29,000 in September, against roughly 90,000 expected, with unemployment at 4.2 percent. July and August were revised down by 31,000 and 29,000. The information sector lost 10,000 jobs, professional and business services lost 9,000, government lost 17,000 and financial activities lost 7,000. Health care added 23,000.
The pattern is the story: the sectors shedding jobs are the white-collar ones where AI tools landed first, and the sectors hiring are the ones where the work still happens with hands. One month proves nothing about cause. A streak of months in the same sectors is harder to wave away, and information has now strung several together. SourcesAB
Weak jobs data sends the Nasdaq to an intraday record and Nvidia past $5.7 trillion
Stocks rallied Friday after the soft September payrolls print cemented expectations that the Federal Reserve holds rates this month. The S&P 500 rose 0.7 percent to 7,722.72 and the Nasdaq Composite climbed 1.2 percent to 27,190.86 after touching an all-time intraday high. Nvidia reached an intraday record of $237.88, putting its market value past $5.7 trillion.
Bad news for workers was good news for the index, because the index is now substantially a bet on AI capital spending, which cheap money extends. The 10,000 people who left information-sector payrolls and the shareholders who gained on the same report are reading the same economy from opposite sides. SourcesBB
Anthropic reportedly plans an October 14th investor day as its listing nears
Anthropic has invited select institutional investors to an October 14th event at its San Francisco headquarters to question senior executives ahead of its Nasdaq listing, Bloomberg reported Thursday, citing people familiar with the matter. Formal marketing is still reported for the week of November 9th, with trading possible before Thanksgiving.
A scheduled investor day is process, not pricing: it establishes that the timetable is being executed, not what the company will be worth. The figures that have leaked so far, including roughly $518 billion in decade-long compute obligations, are what those investors will be asking about. SourcesB
Sean Parker is rebuilding Stability AI around licensed music
Napster co-founder Sean Parker has joined Stability AI's restructuring alongside CEO Prem Akkaraju, refocusing the company on AI music tools for professionals, TechCrunch reported Friday. Stability raised $76 million in late August from Sony, Warner and Universal, which also licensed their catalogs for training. It has shipped three audio models and music-editing software, with melody and beatbox input coming.
Parker's line that he is "playing by the rules this time" is the pitch in one sentence: the labels that sued everyone else are shareholders here. The structural question is whether label-licensed training produces tools musicians prefer, or tools labels prefer. SourcesB
A federal judge calls warrantless Flock searches indiscriminate mass surveillance
Judge Sara Hill of the Northern District of Oklahoma ruled Friday that an officer's warrantless search of Flock's network violated the Fourth Amendment. In United States v. Melisa Kyle, a deputy pulled over a California-plated car and queried a month of its location history from Flock's ALPR database; Hill found the search unsupported by probable cause and called the system "a type of indiscriminate mass surveillance."
The ruling holds that historical location data from always-on camera networks carries a reasonable expectation of privacy even on public roads. Flock's cameras feed AI-driven lookups in thousands of police departments, most of them warrantless today. One district-court suppression order binds no other court, and civil-liberties lawyers will now file its argument everywhere. SourcesB
Huawei ships 100,000 Mate XT 2s as its Tau chip design spreads to the Mate 90
Huawei's consumer chief said the company has shipped about 100,000 Mate XT 2 trifold phones since September 12th and expects total sales to pass 1 million, with launches planned across Asia-Pacific and the Middle East, the South China Morning Post reported Thursday. The phone's Kirin 9050 Pro is the first processor built on Huawei's LogicFolding architecture under its Tau chip-design approach, which stacks logic to shorten signal paths, and the upcoming Mate 90 series will carry Tau-based Kirin chips too.
Tau is Huawei's answer to being locked out of leading-edge fabrication: improve the architecture when you cannot shrink the transistor. A million-unit consumer run is the test of whether that answer ships at volume rather than in keynotes. SourcesB
ESWIN's Hong Kong books close Tuesday with $1.16 billion in cornerstones
ESWIN Computing's Hong Kong offering, open since September 28th, closes October 6th, prices October 7th and is set to list October 9th, which would make it the city's first listed chip company. The deal offers up to 1.57 billion shares at HK$1.48 to HK$1.59 to raise up to about HK$2.5 billion, roughly $320 million, and nine have committed $1.16 billion.
Behind it, Xiaomi-backed robotics-chip designer Zhuhai Amicro has cleared its listing hearing and plans a $100 million-plus Hong Kong with a mid-October debut, the South China Morning Post reported, citing sources. China's chip-listing queue is moving on schedule while the US pipeline waits on Anthropic. SourcesBB
ChatGPT now tries clothes on your photo
OpenAI launched virtual try-on in ChatGPT's shopping results Thursday, globally: upload a selfie or full-body photo, or screenshot an item, and the model renders you wearing it, powered by its Images 2.5 model. A Favorites library for saving products shipped alongside.
This is OpenAI's second run at commerce after its instant-checkout feature underperformed, and it goes straight at Google and Pinterest's try-on features. The business logic is plain: shopping queries are the ad market's most valuable traffic, and a model that holds the shopper through the decision captures what search used to. What OpenAI does with the uploaded body photos is the question its announcement did not answer. SourcesB
An autonomous agent breached the Dutch vulnerability-disclosure institute
The Dutch Institute for Vulnerability Disclosure said an attacker exploited a vulnerability and then turned an automated AI agent loose for post-exploitation inside its network, BleepingComputer reported September 29th. DIVD called the attack "loud and very, very messy," said it could watch the agent working with "sloppy logic," and noted it interfered with its own operations through password spraying. The organization is withholding details during the investigation.
The target was the group that warns everyone else about vulnerabilities, and the attacker still got in with an agent DIVD describes as incompetent. The deployment bar for offensive agents already sits below the skill floor of the humans they replace. SourcesB
Ai2 open-sources the 8-billion-parameter model behind Asta's fast reports
The Allen Institute for AI released AstaBrief 8B, the model behind its Asta research assistant's fast mode, which turns a question plus retrieved literature into a cited scientific report in one pass. Built on Qwen3-8B and trained with and on real user queries, no , it averages 51.1 seconds per report against 178.5 for the Claude-powered thinking mode. Training data ships with it.
An 8B model doing a bounded version of what do in minutes is the economics of the next product generation: route the routine report to the small open model, reserve the expensive one for the questions that resist. SourcesA
MIT and Sakana cut the cost of judging self-improving coding agents
SIFT, from MIT and Sakana AI researchers, uses a language-model judge to rank candidate versions of a self-improving coding agent by pairwise comparison before spending money on full benchmark runs. A comparison costs about 4.4 cents against roughly $6 for a 50-task evaluation. SIFT reached 35.1 percent on the Polyglot benchmark, above the Darwin Godel Machine's 30.7 percent, in under five hours and about $150 of API credits, with a third fewer CPU hours.
Evaluation is the bottleneck in self-improvement loops, and this moves the gate from compute to judgment. The judge is itself a model, so the loop now inherits the judge's blind spots, and nobody has measured those yet. SourcesAB
Chesky says agents need an operating system, not an app store
Airbnb CEO Brian Chesky argued in an interview Thursday that the industry lacks an operating system built for AI: agents should sit at a level with interoperable components and a real SDK rather than competing to be the one assistant. Airbnb shipped AI-powered search this week, plans voice agents this fall, and is wiring the app for agent-to-agent traffic over .
Chesky has exposure: if agents disintermediate marketplaces, Airbnb loses the interface it spent a decade polishing. His OS framing is an argument for a world where platforms stay addressable instead of being scraped by whoever owns the assistant. SourcesB
A paper measures what reinforcement learning costs in solution diversity
"Sharpening Tax in Post-Training" examines 14 base and post-trained model pairs across four families on three agentic benchmarks and finds that RL raises single-shot accuracy while collapsing the diversity that multi-sample inference needs: with light harnesses often beat their post-trained versions on coverage. Post-training pushes tasks toward always-solved or never-solved. The authors propose a Bayesian sampler that tempers sampling per prompt by estimated difficulty and report gains on both metrics.
Labs sell pass@1 because that is what a chat user feels. Anyone running agents that retry should check whether the polished model is quietly the worse one for the job. SourcesA
OneStreamer unifies watching, remembering and deciding when to speak
Nanjing University researchers released OneStreamer, a 4-billion-parameter streaming video model that builds a hierarchical caption memory as it watches, time-anchored details plus event summaries, and learns when to respond proactively rather than only when asked. It reports top results on all eight streaming-video benchmarks it evaluates while supervising 27.5 percent of state tokens, alongside a million-record training dataset.
Streaming video is the modality agents in the physical world actually consume, and a model there has to decide what to keep before it knows whether it matters. SourcesA
Pope Leo XIV: algorithms lack the spark of humanity
Pope Leo XIV wrote Friday that there is "an ontological difference, even before an aesthetic one, between art and what a machine can generate through statistical calculation based on millions of images created by others," adding that "algorithms lack the spark of humanity." It extends the position of his May encyclical on human dignity and automation.
The Vatican is a norm-setter for a billion people, and it has now placed AI-generated art in a different ontological category than human work, before any legislature has managed a definition at all. The copyright lawyers arguing about human authorship just gained an unexpected co-author. SourcesB
Correction, October 4th 2026
The ESWIN item omitted the currency on cornerstone subscriptions. The prospectus states HK$1.161 billion, not US$1.16 billion. The offer timetable is unchanged. SourcesA