Show only each day's banner story

PSI's SI system proves the Aaronson–Ambainis conjecture, open since 2008

SI not only produced a proof of an open mathematical problem but also expressed it in a form a computer checked step by step; this process can be repeated and independently verified.

A research system developed by Physical Superintelligence, using instances of a non-public Anthropic model, obtained a complete proof of the conjecture about quantum speedups on random inputs on its third day of operation. The system then improved the result twice while working on related problems. The conjecture reduces the question of quantum advantage to a property of bounded low-degree polynomials: the proved theorem guarantees the existence of a variable with polynomially large influence. It follows that, in the model under consideration, a quantum algorithm does not provide an exponential advantage on random inputs without additional structure. The Lean 4 formalization has been published alongside the preprint. It contains no unfinished proofs or added axioms; the four final theorems have been rechecked by the Lean kernel and an independent type-checking implementation. Machine verification does not cover the human-readable paper or whether the formal statement matches the original formulation: those remain subjects for human review. David Turturean led the work and is responsible for the paper's correctness, while Jim Halverson audited the proof. Researchers who previously proved a special case are reading the full proof and preparing a shorter exposition.

If the plans materialise, production of key memory for SI accelerators will gain another major player, and Nvidia will gain another way to expand supply.

TSMC considers making 3D DRAM at Nvidia's request

If the plans materialise, production of key memory for SI accelerators will gain another major player, and Nvidia will gain another way to expand supply.

According to The Information, as relayed by Bloomberg, TSMC has started drawing up plans to manufacture three-dimensional DRAM, memory in which cells are stacked in multiple layers to increase density and bandwidth. The initiative comes from Nvidia, TSMC's largest customer, amid a memory shortage for SI systems. Entry by the contract processor manufacturer into this segment could expand its role in the SI hardware supply chain; for now, these are plans under consideration rather than an operating production line.

IPO preparations continue, but the final valuation will test investors' willingness to pay for the growth of a frontier SI laboratory with high computing costs.

Anthropic prepares for its IPO roadshow, but investors consider a $1.5 trillion valuation excessive

IPO preparations continue, but the final valuation will test investors' willingness to pay for the growth of a frontier SI laboratory with high computing costs.

According to CNBC, Anthropic plans to begin its roadshow next week and hold a Nasdaq listing under the ticker ANT in the coming weeks. At the same time, sources speaking to CNBC and Reuters report investor pressure on pricing: one market participant called $1.5 trillion an excessive valuation, while the $1–1.25 trillion range was described as more likely. As recently as September, a valuation of up to $2 trillion was under discussion. According to Reuters, investors are pushing for a lower price because of Anthropic's quarterly losses and substantial commitments to SI-infrastructure partners, and are also waiting for October revenue figures that were not included in the updated Form S-1. Continued volatility could affect the timing of the offering, which some market participants estimate could be as large as $100 billion.

Global markets showed how closely technology company valuations depend on expectations for SI spending, although the partial rebound indicates continuing demand for these stocks.

SI stock selloff spreads to Asia and Europe as Nasdaq begins to rebound

Global markets showed how closely technology company valuations depend on expectations for SI spending, although the partial rebound indicates continuing demand for these stocks.

After Nasdaq fell 5.5% on Thursday, Japan's Nikkei 225 lost 6% on Friday and South Korea's Kospi fell more than 8%; the FTSE 100 was down 1.5% during the day and Germany's Dax fell 2.3%. Investors were assessing a possible slowdown in data center construction after the US midterm elections and the returns on large SI investments. By the US market open, the situation had partly reversed: Nasdaq rose about 1.2–1.8%, and Nvidia gained more than 6%. Additional support came from a report showing 165 thousand new US jobs in October and unemployment falling to 4.1%.

Even after scaling back its initial plans, OpenAI is receiving a substantial capital reserve to develop models and infrastructure, but investors are agreeing to a lower valuation and amount than previously discussed.

OpenAI receives the first $11 billion of a new round at a $1.25 trillion valuation

Even after scaling back its initial plans, OpenAI is receiving a substantial capital reserve to develop models and infrastructure, but investors are agreeing to a lower valuation and amount than previously discussed.

According to Bloomberg, the money has already been transferred to the company, and the remainder of the round is expected to be completed in the coming weeks; the total could reach $20 billion. Coatue Management and T. Rowe Price participated in the first close, while SoftBank, Sequoia Capital, Founders Fund and Andreessen Horowitz were named among the round's other investors. OpenAI initially considered raising $30 billion or more at a $1.4 trillion valuation but scaled back its plans amid the sell-off in SI-sector stocks. The company declined to comment.

New capital will give the laboratory more resources to train models and develop Tinker, but the high valuation raises expectations for future revenue.

Thinking Machines nears $5 billion round at $40 billion pre-money valuation

New capital will give the laboratory more resources to train models and develop Tinker, but the high valuation raises expectations for future revenue.

According to Business Insider, Mira Murati's Thinking Machines is closing a $5 billion round led by Accel at a $40 billion pre-money valuation, and the deal could close as soon as this week. That is more than triple the roughly $12 billion valuation after its $2 billion seed round in 2025. The company develops the Tinker model fine-tuning service and its own Inkling models. A higher valuation was discussed earlier, but those talks did not result in a deal.

Export control for SI chips is shifting from the point of sale to monitoring where they are actually deployed in cloud infrastructure.

Nvidia requires new cloud partners in Southeast Asia to track chip locations

Export control for SI chips is shifting from the point of sale to monitoring where they are actually deployed in cloud infrastructure.

According to Bloomberg, as relayed by iXBT, Southeast Asian cloud providers seeking Nvidia Cloud Partner status must enroll in the Fleet Intelligence program. It uses GPU telemetry and network latency to estimate equipment location and condition. Nvidia intends this to hinder illegal resales of advanced accelerators to China. Outside the new partner requirement, participation remains voluntary, and the system does not automatically send equipment coordinates to the manufacturer.

Researchers now have a reproducible open environment for identifying the causes of robot errors and deliberately changing training data instead of relying on costly manual teleoperation.

Ai2 opens ToyBlocks, a controlled world for training and evaluating robots

Researchers now have a reproducible open environment for identifying the causes of robot errors and deliberately changing training data instead of relying on costly manual teleoperation.

ToyBlocks simulates a robotic arm and blocks on a tabletop and automatically generates tasks, demonstrations and metrics, allowing researchers to study planning, perception and physical control separately. The open package includes code, Molmo 2 and ManiFlow models, 85,336 long-horizon demonstrations, nearly 1,600 hours of execution and 144 million frames. In experiments, skill descriptions written by an agent after its own trials raised the success rate from 24% to 63%. The Molmo 2 and ManiFlow combination, trained only in simulation, completed 50% of tasks on a real Franka FR3 robot; researchers note that further trials are needed for statistical confidence.

The method makes it possible to design a vaccine with future viral variants in mind, but protection in humans has yet to be confirmed.

Cambridge tests SI-designed broad-spectrum bird flu vaccine in animals

The method makes it possible to design a vaccine with future viral variants in mind, but protection in humans has yet to be confirmed.

Researchers collected genetic sequences of bird flu viruses from surveillance programmes worldwide and used computational and evolutionary algorithms to design five synthetic antigens covering the virus's diversity. In a Cell Reports Medicine study, the vaccine fully protected mice and ferrets against a range of genetically distinct variants that had previously infected humans; in the group given the existing WHO candidate vaccine, half of the animals survived. The next step is human clinical trials, and the team plans to apply the same method to seasonal influenza and other rapidly evolving RNA viruses.

Sources 1

The Geneva conference will determine whether states move from general principles of human control to a concrete treaty on autonomous weapons.

Russia and US oppose binding international rules for autonomous weapons

The Geneva conference will determine whether states move from general principles of human control to a concrete treaty on autonomous weapons.

Ahead of the 128-country conference on the Convention on Certain Conventional Weapons in Geneva on 16–20 November, Russia, the US, Israel and several other states oppose a mandate to negotiate a legally binding instrument on a fixed timetable. A 50-country coalition led by Ireland proposes an expert group that would hold three sessions totalling 25 days in 2027 and prepare a draft protocol. According to the ICRC, more than 70 states support opening negotiations. The US considers national restrictions under existing international humanitarian law sufficient; treaty supporters want prohibitions and rules specifically for systems capable of selecting and engaging targets autonomously.

Google has gained direct control over the site, permits and energy strategy for a major project instead of continuing to work through an intermediary developer.

Google buys Crusoe's Wyoming data-center site and will continue the project itself

Google has gained direct control over the site, permits and energy strategy for a major project instead of continuing to work through an intermediary developer.

Crusoe transferred the southeast Wyoming site to Google after completing preliminary work: site selection, obtaining permits and planning the power supply. The previous day, Google announced a $17 billion investment in its first data center in the state, in Laramie County. The Crusoe and Tallgrass Energy project combines fuel cells, grid power and natural gas turbines, with a long-term plan for carbon capture through Tallgrass's regional hub. In 2025, the companies described the possibility of expanding the campus to 10 GW. The terms of the deal were not disclosed. According to Bloomberg, as relayed by The Next Web, Google had previously raised concerns about costs and timelines, while Crusoe had failed to secure customers for the project.

Access to information about advanced SI chips has become a matter not only of export restrictions but also of intelligence and parliamentary scrutiny of industry leaders' personal connections.

US links Megaspeed executive to work for Chinese intelligence to obtain information about Nvidia

Access to information about advanced SI chips has become a matter not only of export restrictions but also of intelligence and parliamentary scrutiny of industry leaders' personal connections.

According to The New York Times, as relayed by The Verge, US intelligence concluded that China's Ministry of State Security recruited Megaspeed executive Alice Huang to gather information about Nvidia CEO Jensen Huang and American SI chips. Under this account, the assignment also included potentially influencing the availability of Nvidia's advanced chips to China. Megaspeed is an Nvidia customer. On November 5, three House committees requested a joint closed-door briefing from Jensen Huang about his contacts with Alice Huang and Nvidia's safeguards against foreign espionage. In the letter, lawmakers asked him to hold it on November 18 and confirm participation or propose another date by November 11.

Buyers of SI computing are beginning to distinguish not only the volume of rented capacity but also the chip generation, price and actual timelines for bringing sites online.

Microsoft winds down its rental of SI computing from Oracle on older Nvidia chips

Buyers of SI computing are beginning to distinguish not only the volume of rented capacity but also the chip generation, price and actual timelines for bringing sites online.

According to Bloomberg, as relayed by The Next Web, Microsoft declined to renew the 2023 agreement and has been switching off the rented capacity in recent months; it was used for purposes including Bing search. Oracle's servers contain previous-generation Nvidia chips and, according to sources, no longer meet Microsoft's needs. This is part of a review of more than $80 billion in commitments to external cloud providers CoreWeave, Nscale and Nebius because of costs and delays. Microsoft declined to comment; Oracle said demand for its SI infrastructure, including older chips, remains strong.

Nvidia has secured a financial interest in Korea's SI ecosystem, while Naver has gained capital to expand its cloud GPU capacity.

Nvidia completes $1 billion investment in Naver and acquires 4.13% stake

Nvidia has secured a financial interest in Korea's SI ecosystem, while Naver has gained capital to expand its cloud GPU capacity.

Naver issued more than 6.5 million shares to Nvidia through a private placement and received 1.3 trillion won, or $1 billion. Nvidia now owns 4.13% of the South Korean company. The investment is tied to the joint SI factory project announced in June: Naver plans to expand GPU infrastructure, starting with its Gak Sejong data center, and grow its GPU-as-a-service computing rental business.

Amazon's SI agent has moved beyond the app and website: purchases can now be completed through a familiar text conversation without a separate interface.

Amazon lets shoppers buy products by texting Alexa

Amazon's SI agent has moved beyond the app and website: purchases can now be completed through a familiar text conversation without a separate interface.

Alexa for Shopping now accepts ordinary text messages: users can describe a desired product by text or photo, receive options, add a selection to their cart and ask the agent to purchase it. The agent also tracks prices, handles recurring purchases and remembers saved preferences. The invitation-only feature is initially available to a small group of US customers, and Amazon plans to expand access in the coming weeks.

Robots can be fine-tuned on precise recreations of situations where they fail instead of waiting for rare cases to occur in the real world.

Physicl raises $10 million to generate training data for rare robot situations

Robots can be fine-tuned on precise recreations of situations where they fail instead of waiting for rare cases to occur in the real world.

Inovia Capital led the Paris-based Physicl seed round with participation from US Venture Partners. Its Studio platform creates synthetic data for specific failures by physical SI models: researchers reproduce a difficult scene and change the lighting, camera angle and object placement. The library contains 10 thousand 3D environments and more than 100 thousand objects. According to the company, early customer SpAItial more than halved the error rate of its room-building model after adding this data. The funds will be used to expand the team and platform.

Banks will gain access to stronger SI defence tools, but the safety of the transition will depend on scrutiny of every new external connection.

South Korea will phase out bank network isolation to enable SI-based defence

Banks will gain access to stronger SI defence tools, but the safety of the transition will depend on scrutiny of every new external connection.

The Financial Services Commission is forming a task force to propose by the end of the first quarter of 2027 a replacement for 2013 rules requiring financial institutions to separate internal networks from the internet. Under the regulator's plan, companies will be able to use cloud and high-performance SI systems to detect attacks, starting with institutions that have more mature defences. Another 13 participants, including NH NongHyup Bank, Kakao Pay and Samsung Life Insurance, have joined the SI cybersecurity pilot. Professor Hwang Suk-jin warns that expanding external connections before equivalent protections are in place will create additional routes for intrusion.

A conversation with an agent can now become a persistent personal application that retains its state and updates data from connected services on its own.

Meta lets users create their own interactive spaces in Muse

A conversation with an agent can now become a persistent personal application that retains its state and updates data from connected services on its own.

A user describes the desired result to Muse, and the assistant creates a persistent space — for example, a trip planner, school dashboard, shopping list or game. It can be expanded and modified through conversation, pinned to the sidebar or main tabs, and scheduled to appear at particular times. Connected external apps allow prices, dates and other data to update in the background, while elements within the space can be controlled directly. The feature is available to all Muse users; Meta is also exploring shared spaces and bringing Active Memory and Active Focus capabilities from Muse Wonder to Muse.

Chemists will be able to screen promising molecules with fewer expensive experiments, even when only a small number of training examples are available for the property of interest.

MIT introduces CheMeleon to predict molecular properties from a small number of examples

Chemists will be able to screen promising molecules with fewer expensive experiments, even when only a small number of training examples are available for the property of interest.

The model is pretrained to predict about 1,600 chemical descriptors for one million molecules from PubChem, then fine-tuned on a small dataset for a specific property — from solubility and viscosity to binding to a protein target. Across 58 test tasks, CheMeleon outperformed other foundation models and classical methods using expert-defined rules. According to MIT, nine of the 20 leading entries in the OpenADMET competition used this model. CheMeleon is available as an open model in the ChemProp package; the paper by researchers from MIT and BASF appeared in the September issue of the Journal of Chemical Information and Modeling.

Sources 1

After the elections, some newly elected officials intend to use permits and rules to restrict construction, but existing projects and the White House position preserve the ability to continue infrastructure work already underway.

US senators and governors-elect renew calls for a pause in data-center construction

After the elections, some newly elected officials intend to use permits and rules to restrict construction, but existing projects and the White House position preserve the ability to continue infrastructure work already underway.

Bernie Sanders called for an immediate nationwide moratorium, while House member Yassamin Ansari proposed putting a pause on advanced SI development and a ban on superintelligence on the new majority's agenda. Texas Governor-elect Gina Hinojosa said she intended to stop new projects until protective rules are adopted; similar demands came from elected officials in Georgia and several members of Congress. In response, White House National Economic Council Director Kevin Hassett told CNBC that construction would continue because local initiatives mostly do not affect existing projects.

The new test measures not impressive answers based on prepared charts, but an agent's ability to check controls and avoid conclusions that go beyond the actual experimental evidence.

OpenAI introduces BioBasicsBench, 100 tasks analyzing real laboratory data

The new test measures not impressive answers based on prepared charts, but an agent's ability to check controls and avoid conclusions that go beyond the actual experimental evidence.

According to OpenAI's description, as relayed by Unite.AI, the benchmark was developed with Ligo and tests the interpretation of raw data from routine experiments: gel and microscopy images, sequencing reads, instrument tables, photographs and protocol videos. The tasks cover protein methods, imaging, cytometry, quantitative assays, nucleic-acid methods and molecular cloning. An agent must take measurements, detect errors and recognize cases where the evidence does not support a conclusion. Seven models were tested in 37 configurations, with five attempts per task — 18,500 graded responses in total. At the highest shared reasoning setting, GPT-6 Astra received 66.7% of the available credit, compared with 43.8% for GPT-5.6 Sol. OpenAI also cites errors made by Astra: selecting the wrong lanes in colony PCR, contradicting its own observations in a conclusion about cellular vacuoles, and overcounting colonies by about 12%. The result shows improved analysis but not reliable independent laboratory judgment.

Working agents rather than individual questions from people already generate most of the platform's traffic, while caching and competition make each million tokens significantly cheaper.

OpenRouter handled 790 trillion tokens in October, with agents generating 76%

Working agents rather than individual questions from people already generate most of the platform's traffic, while caching and competition make each million tokens significantly cheaper.

Monthly request volume through OpenRouter grew 37% from September and almost 25-fold from January. Agentic API keys accounted for about 600 trillion of October's 790 trillion input and output tokens; their share was 8% a year earlier. OpenRouter classifies keys using seven signals, including tool calls, number of interactions and intervals between them. Repeated submission of long context means that the share of tokens is not equal to the share of unique text. The median share of agentic input tokens served from cache by the providers included in the sample rose from 76.1% in June to over 90% in October. Since January, the effective price per million tokens fell from $1.21 to $0.52 for closed-weight models and from $0.28 to $0.084 for open-weight models. After temporarily falling to 52%, open weights' share of paid tokens again exceeded 70%, primarily thanks to DeepSeek V4.1 Flash.

Companies are using more SI for less money. That makes adoption easier but prevents labs from maintaining previous revenue growth through request volume alone.

Companies' spending on SI tokens grew just 1.5% across August and September

Companies are using more SI for less money. That makes adoption easier but prevents labs from maintaining previous revenue growth through request volume alone.

According to Ramp, combined token spending grew 1.5% across August and September, compared with 42% in July alone. Economist Ara Kharazian links the slowdown not to declining use but to cheaper models, discounts, routing to less expensive options and customers' own spending limits. The median company spent $12–15 per employee per month, the top 10% spent $600–1000, and spending by the most active 1% of companies edged down from its summer level of just under $10 thousand per employee. Spending rose again by 8.9% from September 23 to October 23, but Kharazian considers the previous growth rates vulnerable: according to his data, nearly all demand comes from fast-growing, venture-backed technology companies. Growth in token volume therefore does not guarantee proportional revenue growth for model providers.

Open weights allow a model's political framing to be changed substantially with relatively little fine-tuning while preserving its general capabilities; for now, however, only the developer's own methodology confirms the result.

Nous Research fine-tunes Qwen3.8-27B and cuts answers echoing Chinese state rhetoric from 22% to 0.9%

Open weights allow a model's political framing to be changed substantially with relatively little fine-tuning while preserving its general capabilities; for now, however, only the developer's own methodology confirms the result.

Nous Research released the weights of Arete-27B under Apache 2.0, a version of Alibaba Qwen3.8-27B fine-tuned on answers from the Inkling model. Training comprised two passes over 28,886 examples, about 52.8 million tokens; reinforcement learning, preference optimization and architecture changes were not used. On Nous's own set of 803 prompts about freedom of expression, ideology, economics, education, privacy, democracy and the environment, the share of answers in which the model reproduced or endorsed China's official position fell from 22% for base Qwen to 0.9% for Arete with reasoning disabled. The average score for conveying specified facts rose from 2.32 to 2.98 out of 5. Grok 4.7 evaluated answers without being told which model wrote them; the methodology, data and evaluation code are public, but the results have not undergone independent review. On the internal Hermes Bench, the result remained almost unchanged at 65.60%, versus 65.77% for the base model.

Trip planning could become a ready-made set of interactive options and bookings rather than a sequence of text answers.

Airbnb will launch an SI travel agent in the first half of 2027

Trip planning could become a ready-made set of interactive options and bookings rather than a sequence of text answers.

Airbnb is developing an agent that, instead of relying on a single chat window, will assemble an interface for each request from maps, photos, comparisons, calendars and itineraries. It will help users choose and book accommodation, discuss options with fellow travellers and eventually combine homes, hotels, services and experiences in one trip plan. The launch timing was disclosed in the third-quarter shareholder letter: the first half of 2027.

Tesla is shifting production capacity from a slow-selling pickup to robotaxis, making autonomous transport a higher priority.

Tesla will wind down Cybertruck production in Texas to make room for Cybercab robotaxis

Tesla is shifting production capacity from a slow-selling pickup to robotaxis, making autonomous transport a higher priority.

Elon Musk said Tesla will begin winding down Cybertruck production at its Texas Gigafactory around mid-2027 to free capacity for the two-seat self-driving Cybercab. He did not specify whether the pickup would be produced at another site. Cox Automotive estimates that 20,237 Cybertrucks were sold in the US in 2025, down 48% from the previous year. The Cybertruck is made in Texas alongside the Model Y and Cybercab.

Speech generation is becoming available directly on personal devices: these features can work offline without sending voice data to an external service.

Underdog releases Woof, an open 32 MB speech model

Speech generation is becoming available directly on personal devices: these features can work offline without sending voice data to an external service.

The model has 31 million parameters and, according to Underdog, runs locally on most devices without an internet connection. A browser demo in the form of a game speaks text entered on a virtual typewriter and generates music. Woof's weights are available for download. The company previously released the 167 MB Bark speech recognition model; both projects are intended for local assistants and voice processing without sending data to the cloud.

Opposition to new computing capacity has escalated into direct intervention at a construction site, making energy use and urban restrictions part of SI infrastructure timelines.

Protesters enter Microsoft data center construction site in Amsterdam

Opposition to new computing capacity has escalated into direct intervention at a construction site, making energy use and urban restrictions part of SI infrastructure timelines.

About 100 demonstrators entered the site in the Sloterdijk district, and some climbed a construction crane. Police ended the demonstration on the mayor's orders, and several arrests were reported. Geef Tegengas and Fossielvrij NL organised the protest against three 85-metre towers that Pure Data Centres is building for Microsoft in an investment exceeding $1 billion. Expected electricity consumption and the project's impact on housing plans were among the stated concerns. The project received approval before the city's current restriction on new data centers through 2030.

Government plans now rely on a more cautious forecast: the benefits of SI remain substantial, but their scale depends directly on the speed of adoption.

UK government estimates SI could add £55–140 billion to the economy in 2030

Government plans now rely on a more cautious forecast: the benefits of SI remain substantial, but their scale depends directly on the speed of adoption.

SI minister Kanishka Narayan disclosed the government's own estimate: SI adoption and investment in data could add 0.4–1.3 percentage points to productivity growth, equivalent to £55–140 billion in gross value added in 2030, with a central scenario of £120 billion. The AI Opportunities Action Plan previously used an estimate of more than £400 billion from a 2023 Public First report commissioned by Google. Narayan stressed the high uncertainty of the calculations, while Public First explained that its model assumed faster adoption of the technology.

Environmental services will be able to identify suspicious sites from satellites with fewer errors and spend less time on manual checks.

PNRPU raises accuracy of detecting illegal dumps in satellite images to 95%

Environmental services will be able to identify suspicious sites from satellites with fewer errors and spend less time on manual checks.

Researchers at Perm National Research Polytechnic University supplemented a training set of more than 100 thousand satellite images with quarries, industrial zones and construction sites that visually resemble dumps. A separate program automatically added characteristics of the surrounding area, including land cover, population density and terrain. According to the team, classification accuracy rose from 56% to 95%, while false positives fell roughly 24-fold. The results were published in the journal Applied Mathematics and Control Questions.

Drone groups will be able to maintain positioning when satellite signals are lost and coordinate their actions in one local coordinate system.

Wheelies combines three navigation methods for autonomous drone groups

Drone groups will be able to maintain positioning when satellite signals are lost and coordinate their actions in one local coordinate system.

The Russian company's onboard module combines GNSS satellite navigation, visual odometry from successive camera frames and ultrawideband ground radio beacons. The system checks the reliability of satellite coordinates and automatically switches to visual motion estimation when they are lost; the stated distance error is up to 1% of the route. Beacons give multiple aircraft a shared coordinate system for coordinated operations and precision landing. A demonstrator has been built, and Wheelies is using it to develop data exchange, collision avoidance and local group coordination.

Sources 1

The case has moved to judicial proceedings, where prosecutors must prove the installation of the recording device and the surrounding circumstances.

OpenAI employee pleads not guilty in office listening-device case

The case has moved to judicial proceedings, where prosecutors must prove the installation of the recording device and the surrounding circumstances.

Abdulrahman Alfozan, a 35-year-old OpenAI engineer and venture partner at Raed Ventures, pleaded not guilty at a hearing in federal court in San Francisco. A grand jury charged him with interception of oral communications: according to prosecutors, on September 28 he placed a recording device inside a television in an executive conference room at OpenAI. Alfozan has remained in custody without bail since early October. The next hearing is scheduled for November 17; if convicted, he faces up to five years in prison.

Sources 1

Alphabet has acquired a financial stake in a manufacturer expanding from smartphones into SI infrastructure; the filing alone does not signify a new technology partnership.

Alphabet discloses a $548 million stake in Qualcomm

Alphabet has acquired a financial stake in a manufacturer expanding from smartphones into SI infrastructure; the filing alone does not signify a new technology partnership.

In its quarterly Form 13F, Alphabet reported 2,976,982 Qualcomm shares worth $547.9 million as of September 30, 2026. The position was established in the third quarter. Qualcomm is developing processors for SI data centers and also remains Google's longtime Android partner, but the filing itself does not connect the investment to that line of business.

Protecting SI agents increasingly depends on testing the entire chain of connected services and permissions, not just the behavior of an individual model.

OpenAI brings in Hacktron to find vulnerabilities in connections with third-party services

Protecting SI agents increasingly depends on testing the entire chain of connected services and permissions, not just the behavior of an individual model.

According to OpenAI's description, as relayed by Unite.AI, company engineers and Hacktron researchers will jointly select priority areas for testing, focusing on weaknesses that arise when applications, identity systems and infrastructure are connected. Hacktron will receive expanded access to frontier models through the Daybreak Red program for authorized vulnerability discovery and security testing. The companies have previously interacted through OpenAI's vulnerability bounty program.

Microsoft is losing a dedicated group that built agent tools and explored new architectures; Langford links his departure to scaling taking priority over research.

John Langford leaves Microsoft and declares its AI Frontiers research group finished

Microsoft is losing a dedicated group that built agent tools and explored new architectures; Langford links his departure to scaling taking priority over research.

Researcher John Langford said AI Frontiers, which he founded with colleagues inside Microsoft Research in 2023, has stopped operating after being moved into Microsoft AI in early September. The group had about 50 people; its projects included the Autogen multi-agent framework and the Fara computer-use agent. Langford writes that he left Microsoft over a mismatch of approaches: in his assessment, Microsoft AI leadership prioritizes scaling models rather than researching new architectures. He names Next Latent, Dion3 and Full Bandwidth Transformer among the directions that failed to gain traction. This is Langford's own explanation; his post contains no response from Microsoft to that assessment.

Sources 1

An agent's long-term memory requires more than a large context: the system must select useful experience, correct its notes and prepare them for future tasks.

Anthropic breaks agent memory into notes, salience and log processing between sessions

An agent's long-term memory requires more than a large context: the system must select useful experience, correct its notes and prepare them for future tasks.

In Anthropic's engineering analysis, agent memory means knowledge from past interactions and assignments retained as context available in later sessions, rather than changes to model weights or compression of the current dialogue. Claude can mark important events as they happen, write them to files and later use those notes. A separate mechanism called dreaming processes logs between interactions: it corrects mistakes, creates notes and new memory representations, and selects information to load into future sessions. Anthropic compares this with memory consolidation during sleep and uses Continual Learning Bench tasks to show how an agent can formulate a rule after an error and apply it to the next question.

Healthcare teams with limited budgets will gain direct access to computing and agentic services for diagnostics, research and organizing care.

AWS commits $100 million in cloud and SI resources to healthcare projects

Healthcare teams with limited budgets will gain direct access to computing and agentic services for diagnostics, research and organizing care.

Amazon Web Services has launched the five-year, $100 million Health Impact Initiative. Nonprofits, government agencies, research institutions and technology companies will be able to receive cloud computing, SI and agentic services for projects serving underserved populations. Support covers both prototypes and deployment of working solutions. Priorities are access to healthcare, research into women's and children's health, and food-related projects. Applications will be reviewed weekly; starter credits will be up to $10 thousand, with further amounts calculated according to project needs. The initiative continues AWS programs that have supported more than 600 healthcare organizations in over 80 countries since 2021.