Show only each day's banner story

SI agents breached more than 100 company websites, about $50 per successful break-in, with John Deere, SharkNinja and Cummins among the targets

Hacking has stopped being expensive: agents do it for the price of lunch per site, so defenders will have to work just as fast and at the same scale.

According to a report by NorForge researcher Snorre Fagerland, a Chinese-speaking hacker ran agents on the open-source Cairn platform starting September 25: of more than 1,200 attempts, over 100 ended in a break-in, at about $50 in tokens per successful breach and roughly $5,000 for the whole campaign. The agents tried to plant web shells (programs for remote command execution) and gain a foothold in shop and payment systems. Hundreds of credentials and API keys were found on open control panels. Reuters confirmed the targets included John Deere, SharkNinja and Cummins: Deere said no sensitive data was affected, SharkNinja said the attack hit a product support site, and Cummins is investigating a possible compromise of a vendor's site. Judging by the name of the worker process, "claudecode_deepseek-v4.1-flash", TNW suggests Claude Code and the DeepSeek V4.1 Flash model were used; this is an inference from the name.

Sources 1

thenextweb.com

US government services will stop using DeepSeek models; other users can still access them as before.

The US bars civilian agencies from using DeepSeek and buying its products

US government services will stop using DeepSeek models; other users can still access them as before.

The US Department of Homeland Security ordered DeepSeek products and third-party services built on its models removed from the IT systems of civilian federal agencies, and excluded the company from government procurement and contracts. According to The Next Web, the order signed by Secretary Markwayne Mullin is dated July 24 and relies on the 2018 supply chain security law; the notice reached the Federal Register on October 19. The department cited "multiple threat vectors": possible transfer of sensitive data to China, the possibility of producing "actionable intelligence" on US persons and agencies, and possible malware and vulnerabilities. The order does not cover military agencies. DeepSeek and the Chinese embassy in Washington had not commented at the time of the reports; DeepSeek's position was not found in the sources checked.

Sources 3

thenextweb.comvedomosti.rubnnbloomberg.ca

The first such proof has been checked by a machine, and mathematicians say SI has reached the level of significant theorems, even if in a weakened form.

Mathematicians assess an SI proof of a weakened Riemann hypothesis: the 199-page version is checked in Lean, the short one is not yet

The first such proof has been checked by a machine, and mathematicians say SI has reached the level of significant theorems, even if in a weakened form.

According to Quanta, OpenAI published the result on October 6 among 377 mathematical papers: the "quasi-Riemann hypothesis", a weakened form of the statement about the distribution of prime numbers. The first version (September 30, zeros in the strip from 1/8 to 7/8, 199 pages) was formally verified in the Lean system; the second (October 5, from 1/12 to 11/12, 49 pages) has not yet been formally verified but, in experts' assessment, appears correct. Jared Duker Lichtman called it "the most significant SI-for-math achievement so far, by quite a margin" and Maksym Radziwiłł called the implications "staggering", but noted that in presentation the papers would be "immediately rejected at any mathematical journal". The proof combined known methods rather than introducing new concepts.

Sources 1

quantamagazine.org

The US government signals that it is ready to use export controls to get the behavior it wants from SI labs.

US deputy defense secretary says export controls are a legitimate lever against SI labs that do not cooperate with the government

The US government signals that it is ready to use export controls to get the behavior it wants from SI labs.

According to The Next Web's account of an interview with The Information, Emil Michael, the Pentagon's under secretary for research and engineering and vice chair of the White House working group on "superintelligence", called export controls a suitable way to get SI companies to meet government requirements, citing past limits on encryption exports and on foreign nationals' access to technology. He alleged that when Anthropic refused to "pull back a model" the government considered jailbroken, export controls were applied to the company "within two hours". He also said Anthropic remains excluded from Pentagon contracts as a "supply chain risk" and that a return is not being discussed unless the company changes its approach to "constitutional SI". Anthropic's position on these claims was not found in the sources checked.

Sources 1

thenextweb.com

If the claims hold up, a fast and cheap model for agents will also become more accessible in open form.

MiniMax releases M3.1 Flash for agents: claimed to be three times faster than its predecessor, weights promised for October 23

If the claims hold up, a fast and cheap model for agents will also become more accessible in open form.

According to MiniMax, the model's inference speed rose almost threefold and its cost in agent scenarios fell by 70%; generation reaches up to 200 tokens per second. The API costs $0.15 per million input tokens and $0.60 per million output tokens. Before the announcement the model ran anonymously under the name "Space Bunny" and, in the company's words, processed more than 200 trillion tokens. Results on AutomationBench, DeepSWE v1.1, Agents' Last Exam and BabyVision were published by MiniMax itself. The company set the release of the weights for October 23.

Sources 1

habr.com

Training models on Amazon's chips gets easier: no need to rewrite the program, so developers have more choice than just Nvidia.

Amazon releases TorchNeuron: PyTorch code runs on Trainium chips almost without rewriting

Training models on Amazon's chips gets easier: no need to rewrite the program, so developers have more choice than just Nvidia.

TorchNeuron is a PyTorch backend for AWS Trainium2 and Trainium3 chips under the Apache 2.0 license. According to Amazon, "a single configuration change" is enough to move code: in a demonstration, the same training script ran on conventional chips and on Trainium, differing only in the hardware setting. It is compatible with TorchTitan and FSDP (distributed training) and with existing PyTorch testing and debugging tools. The company names Cartesia and the Lieber Institute for Brain Development among early users; publication on GitHub is announced for October 20, at the PyTorch conference.

Sources 1

aboutamazon.com

An agent with no assignments started inventing activities for itself, and they helped it later, but what that means is not yet clear.

MIT researchers left copies of Claude on a virtual island for 30 hours: the agents invented games and then handled tasks better

An agent with no assignments started inventing activities for itself, and they helped it later, but what that means is not yet clear.

A team at MIT CSAIL led by postdoc Antonio Norelli placed single copies of Claude Opus 4.7 on a virtual island on the Roblox platform: the agent could move an avatar, build blocks and keep text notes. Using the criteria of psychologist Gordon Burghardt, the researchers found "play": their own rules and challenges, repeating actions with variations, using objects for something other than their purpose, and goals without external reward. Agents with island experience did better, for example, at stacking objects on the highest point; deleting memory made results worse. The authors admit it is unclear whether the agents were actually playing or only pretending to play; cognitive scientist Eunice Yiu noted that the paper cannot settle this. The work has not been peer reviewed.

Sources 1

technologyreview.com

The military wants to hand part of its cyberattacks to autonomous SI agents, and it is unclear how it will limit the damage.

US Army Cyber Command is developing ECHOTRIBBLE, autonomous SI agents for attacks on adversary computer systems

The military wants to hand part of its cyberattacks to autonomous SI agents, and it is unclear how it will limit the damage.

According to documents obtained by The Intercept (a request to contractors and answers to their questions), US Army Cyber Command (ARCYBER) is building a platform meant to "identify and neutralize identified threats/targets," primarily in offensive operations. According to the documents, the software must operate "with minimal human oversight" and act on its own on predefined triggers; "neutralizing" can mean disabling a threat completely or temporarily, or deleting it. The agents are to spread across networks, including those disconnected from the internet, and afterward delete themselves, ideally unnoticed. The article gives no timeline or budget. David Krueger, a professor at the University of Montreal, called building autonomous systems with a mandate to "neutralize threats" "incredibly dangerous"; Hyrum Anderson, head of the company Abundant, asked whether an agent might pass through a US water utility's network on its way. ARCYBER did not respond to the outlet's questions, including those about measures to prevent accidental damage.

Sources 1

theintercept.com

A fast new type of model has appeared that answers almost instantly and costs little, if the stated figures hold up.

Inception releases Mercury 3, a diffusion language model running at over 1,000 tokens per second

A fast new type of model has appeared that answers almost instantly and costs little, if the stated figures hold up.

A diffusion model does not write text token by token but refines many tokens in parallel. According to Inception's own tests, Mercury 3 outperforms GPT-6 Luna, Claude Haiku 5.5 and GLM-5.3 Flash while running 4-8 times faster: 1,068 tokens per second on Nvidia GPUs, with a 1-million-token context (previously 260,000). On 50 live Mind2Web web-navigation tasks the model completed 50%. The price is $0.20 per million input tokens and $0.70 per million output tokens, half that at launch; access is through the Inception API, OpenRouter, Baseten and Vercel. The Next Web notes that the data is claimed by the company itself, and there are no independent tests yet.

Sources 2

inceptionlabs.aithenextweb.com

Companies now have an official template for showing that their quality control of SI meets EU law.

The first harmonised standard for the EU AI Act, on quality management systems, is published in the Official Journal

Companies now have an official template for showing that their quality control of SI meets EU law.

The standard EN 18286:2026, a quality management system for the purposes of the EU AI Act, developed by CEN and CENELEC, was published on October 19 by Commission decision 2026/2274. It sets out the requirements of Article 17 of the Act for providers of high-risk systems as a working quality management system across the whole lifecycle, and is designed to be combined with ISO/IEC 42001 and ISO 9001. Providers who apply the standard get a presumption of conformity with the Article 17 requirements; this is the only such standard on SI in EU law.

Sources 1

digital-strategy.ec.europa.eu

Talks on rules for combat SI systems continue: Russia, by its own account, is against an immediate ban.

Russia's Foreign Ministry: Russia does not support an immediate ban on lethal autonomous weapons systems

Talks on rules for combat SI systems continue: Russia, by its own account, is against an immediate ban.

According to Oleg Postnikov, director of the Russian Foreign Ministry's Department for Non-Proliferation and Arms Control, Russia does not support the efforts of "radically-minded states" pushing for an immediate ban on such systems. He argued that existing international law, including humanitarian law, already contains a broad set of restrictions, and that states first need to analyze the existing norms and agree on common terminology. Postnikov said the US also rejects limits on autonomous weapons. The statement was made ahead of the review conference of the Convention on Certain Conventional Weapons on November 16-20, where a group of governmental experts is working. The position of those who favor a ban was not found in the sources checked.

Sources 1

interfax.ru

Telecom operators' data centers are among the facilities that parties report striking, so backup sites become a necessary part of the service.

Operator Datagroup reports its Kyiv data center was completely destroyed in a strike on October 19; Russia's Defense Ministry named it as a target

Telecom operators' data centers are among the facilities that parties report striking, so backup sites become a necessary part of the service.

According to Datagroup CEO Mykhailo Shelemba, the data center's infrastructure in Kyiv was completely destroyed in an overnight ballistic strike and restoration is impossible. According to the company, no employees were hurt, customer data is intact, and services continue from backup sites. According to Russia's Defense Ministry, in the course of the special military operation, a "data processing center, the Data Group data center", the Fire Point logistics hub and the Khimvolokno enterprise were struck in Kyiv. According to Interfax-Ukraine, this is the second strike on the data center: the company recorded damage from an earlier strike at the end of September.

Sources 2

en.interfax.com.uainterfax.ru

The court will still decide whether to keep the suspect in custody; the prosecution links the case to contacts in Saudi Arabia, and the defense disputes this.

US prosecutors ask to hold an OpenAI employee without bail in the bugging case, citing messages with a Saudi prince

The court will still decide whether to keep the suspect in custody; the prosecution links the case to contacts in Saudi Arabia, and the defense disputes this.

According to the prosecution's materials, 35-year-old Saudi citizen Abdulrahman Alfozan, arrested on October 9, installed a recording device in OpenAI's executive conference room in San Francisco on September 28. Prosecutors say that in 2025-2026 he messaged frequently on WhatsApp with an unnamed member of the Saudi royal family, who invited him to join a Saudi SI initiative and asked when he would leave OpenAI; in one message he wrote: "Saudi needs you more than the U.S.!" According to the prosecution, after the arrest a Saudi military member bought tickets to Doha for the defendant's family. Lawyer William Pernik said the trip had been planned for family and immigration reasons and that nationality does not mean a flight risk. No espionage or foreign-agent charges have been filed; the maximum sentence is five years.

Sources 1

ynetnews.com

Companies get a way to give agents exactly the permissions a task needs, and to keep watch over them.

Microsoft adds SI agent controls to Microsoft 365 E7: zero access by default, separate identities, supervisor agents

Companies get a way to give agents exactly the permissions a task needs, and to keep watch over them.

According to Microsoft, an agent gets zero access, and only for the duration of a specific task (dynamic permissions); risk is assessed continuously from signals about the agent's identity, data, devices, network and behavior. Entra Conditional Access for agents can, for example, remove the right to change files while keeping read access, require human confirmation, or stop the agent. The definition of agents has been expanded by 35 types (60 in total), including Claude Code, Codex and GitHub Copilot; each agent is assigned its own Entra Agent ID; "guardian agents" (Guardian Agents) in Project Perception, the company says, "watch every action of every agent before it is executed." TechCrunch recalls that Project Perception was introduced in June as a response to hacking models, built on the assumption of attacks from outside, and that Microsoft now warns the agents companies use themselves can also pose a danger. The announcement gives no dates or prices.

Sources 3

thenextweb.comblogs.microsoft.comtechcrunch.com

Nvidia has a growing number of rivals getting funded to build SI servers.

SI chipmaker Tenstorrent raises $1.5 billion at a $6.2 billion valuation

Nvidia has a growing number of rivals getting funded to build SI servers.

The $1.5 billion Series E round was led by Trace Capital and Innovation X; investors include ARK Invest, Fidelity, Morgan Stanley, T. Rowe Price, SanDisk and Tether Investments (according to techstartups), and TechCrunch also names Cadence Ventures and Eclipse. The post-round valuation is $6.2 billion, more than double the $2.6 billion of late 2024. Five months ago, Intel and Qualcomm were reportedly interested in buying the company at a valuation of up to $5 billion. Jim Keller's company develops SI processors, Galaxy servers, processors on the open RISC-V architecture and licensable blocks; Keller told Bloomberg that Tenstorrent decided to stay independent after showing that Galaxy servers can run DeepSeek models. The money will go toward chips for serving open-weight models and video generation.

Sources 2

techstartups.comtechcrunch.com

Anthropic, judging by the report, is discussing another large compute order, but the deal has not been signed.

Nscale and Anthropic are in talks on a new $45 billion compute agreement

Anthropic, judging by the report, is discussing another large compute order, but the deal has not been signed.

Bloomberg reported the talks, citing people familiar with the matter; TechCrunch calls it "yet another" $45 billion deal, and details may change. Nscale and Anthropic did not respond to TechCrunch's request for comment. Both companies are preparing to go public.

Sources 1

techcrunch.com

Germany and Nvidia agreed to develop SI for factories together, and Nvidia is investing in German startups.

Germany signs a declaration of intent with Nvidia on SI for industry; Huang announces investments in four German startups

Germany and Nvidia agreed to develop SI for factories together, and Nvidia is investing in German startups.

At the Invest in Germany summit, Chancellor Merz and the Federal Ministries for Economic Affairs and for Research signed a joint declaration with Nvidia: the parties intend to develop SI models for the physical world and expand cooperation with the Industry 4.0 Platform. Huang announced planned investments in Agile Robots, Avelios Medical, Mirelo and Reflex Aerospace; according to Nvidia, this is part of its commitments in Europe of nearly $10 billion. Euronews reports that last week in London Huang warned that Germany is falling behind in the SI race, and that the government is preparing to commit hundreds of millions of euros, mainly to buying compute in Deutsche Telekom's Industrial AI Cloud on Blackwell chips.

Sources 2

blogs.nvidia.comeuronews.com

Quantum processors can be tested on a supercomputer before they are built, but the record has so far been claimed only by Nvidia.

Nvidia reports a record simulation of a 104-qubit fault-tolerant quantum processor

Quantum processors can be tested on a supercomputer before they are built, but the record has so far been claimed only by Nvidia.

According to Nvidia, its researchers and the Jülich Supercomputing Centre simulated a 104-qubit fault-tolerant quantum processor on Europe's exascale supercomputer JUPITER (GH200 chips); Nvidia itself calls it the largest full simulation of such a processor. They modeled "magic state cultivation", an approach that, the company says, cuts the number of physical qubits needed by 90% compared with earlier methods. Development was done on a GB200 NVL72 system.

Sources 1

nvidianews.nvidia.com

A selfie may become a source of medical assessment, but the model is trained on a narrow group so far and needs testing on other people.

SI model estimates cancer patients' survival odds from a face photo, study in Lancet Digital Health

A selfie may become a source of medical assessment, but the model is trained on a narrow group so far and needs testing on other people.

The FAHR-Face model, described by radiation oncologist Ray Mak of Mass General Brigham, was trained on 40 million unlabeled face photographs, then tuned to estimate biological age and survival odds in cancer. It was tested on 3,820 patients and additionally on a group of more than 38,000; in the smaller group, of 122 patients with the highest risk, 64 (52%) died within 90 days of the photo. The training sample is mostly patients from the Boston area, predominantly white. Bioethicist Matthew Liao (New York University) warned that algorithm errors could deprive someone in need of care, and about the danger of use by insurers and employers.

Sources 1

scientificamerican.com

Training frontier models already requires dedicated gigawatt-scale sites, and the key question becomes the reliability of such systems.

Meta details its Prometheus supercluster of about a gigawatt in Ohio, on which it trained the Muse models

Training frontier models already requires dedicated gigawatt-scale sites, and the key question becomes the reliability of such systems.

According to Meta's description, the New Albany (Ohio) site, which two years ago served web traffic, has been rebuilt for training frontier models. The first phase is complete and was used to train the Muse models, including Muse Wonder; construction is proceeding in phases. Meta reports that in training Muse Wonder it reached more than 85% goodput in pretraining and more than 90% in reinforcement learning; the mean time between interruptions was about 47 thousand GPU-hours, and after firmware fixes more than 54 thousand. All figures are the company's own data.

Sources 1

engineering.fb.com

SI is learning to move ancient banking programs to modern languages, but without human reviewers it is not yet trusted.

OpenAI reports a marked improvement in its models' quality on COBOL, Fortran and other "rare" programming languages

SI is learning to move ancient banking programs to modern languages, but without human reviewers it is not yet trusted.

The facts below are from Engadget's account. OpenAI built the MultiLangBench set, where a model must write an SQLite-compatible database engine in a given language; GPT-6 Astra passed 87.4% of tests on average across 20 languages (GPT-5.6 Sol, 26.8%), with COBOL the hardest at 72%. When migrating programs from COBOL to C++, Java, Kotlin, Python or Rust while preserving behavior, GPT-6.1 Sol showed 97.9% accuracy at high reasoning effort and 86.6% at low. According to OpenAI, a migration counts as successful only if all tests pass, and production systems need expert review. Engadget cites OpenAI's estimate that about 40% of the world's banks still use COBOL.

Sources 2

openai.comengadget.com

For now this is a promise.

Musk promises to give one SpaceX share to every child in the US starting next year

For now this is a promise.

Musk wrote on X: "Next year, I will give one of my SpaceX shares to every kid in America. This will be the first of many." By NBC's calculation, that means about 72 million children under 18 and shares at $150, or roughly $10.8 billion, less than 2% of Musk's net worth. He did not disclose the mechanism, timing or conditions of the handover. The statement came a day after Anthropic announced it would commit 5% of its shares to a public benefit fund; Musk suggested Anthropic distribute shares "directly to the public". NBC notes that Musk's earlier big announcements were not always carried out.

Sources 1

nbcnews.com

Workplace SI assistants are starting to interact with each other, and investors are ready to pay billion-dollar valuations for it.

Startup Town raises $108 million at a $1 billion valuation for SI assistants that work together inside a team

Workplace SI assistants are starting to interact with each other, and investors are ready to pay billion-dollar valuations for it.

The Series B round was led by Forerunner Ventures and Menlo Ventures, with a16z and First Round Capital participating. Town gives each employee an assistant, a "Townie", connected to email, documents and calendar; the assistants can now ask each other questions and hand off work, while the person controls what is shared. According to the company, annual recurring revenue passed seven figures within weeks of the June launch, the number of paying customers has quadrupled since July, and customer data is not used to train models.

Sources 1

thenextweb.com

Chatbots' hidden rules are sometimes visible in page code, and they show what the bot has been told not to say.

About 80 WhatsApp help pages exposed hidden instructions for a Meta chatbot, including a ban on advising people to contact oversight bodies

Chatbots' hidden rules are sometimes visible in page code, and they show what the bot has been told not to say.

According to an analysis by the company Redact, the code of about 80 WhatsApp help pages contained roughly 100 blocks of text written as instructions for a Meta language model: for example, sections titled "WHAT NOT TO SAY" and a ban on giving legal advice. On the page about deleting an account, among other things, there was an instruction not to recommend contacting a data protection authority or another regulator. Wired saw this text on the pages. Meta said the cause was a bug.

Sources 1

wired.com

To tell people from deepfakes, video calls are starting to require an iris-based check.

Zoom and financial firms adopt a "proof of human" badge from Altman's World against deepfakes

To tell people from deepfakes, video calls are starting to require an iris-based check.

World, which according to Fortune is co-founded by Sam Altman, unveiled an identity verification system: in Zoom, a verified participant gets a blue human badge. Financial firms adopting the system include Bain Capital and VanEck Funds; they will keep iris-scanning "orb" devices in their offices to enroll employees. Verification is available at three levels: "Proof of Attributes" (government documents), "Proof of Presence" (a selfie) and "Proof of Human" (the app, checked against the orb's biometric records).

Sources 1

fortune.com

Greece now has its own computing for SI, and startups in the region will be able to use it for free from 2027.

Greece launches the Daedalus supercomputer: 85.7 petaflops on Nvidia GH200 chips

Greece now has its own computing for SI, and startups in the region will be able to use it for free from 2027.

Daedalus in Lavrio (Attica), in a former power station building, delivers 85.69 petaflops; in the June 2026 Green500 ranking it is 31st in the world by power and 23rd by energy efficiency. The liquid-cooled system is built on Nvidia Grace Hopper GH200; it cost EUR 58.9 million, with HPE supplying the equipment (EUR 41 million) and Mainsys the installation (EUR 17.9 million), funded by the EU recovery fund and EuroHPC. Through the Pharos AI Factory, commercial services start in March 2027: free for startups and small businesses within state aid limits. A second supercomputer is planned in Kozani.

Sources 1

euronews.com

If the plan is right, Waymo's robotaxi fleet will grow fivefold in two years and become more visible on the streets.

Waymo plans to grow its robotaxi fleet to 20,000 vehicles by the end of 2027

If the plan is right, Waymo's robotaxi fleet will grow fivefold in two years and become more visible on the streets.

According to Semafor, citing people familiar with the plans, Waymo now has about 4,000 vehicles, should have 8,000 by the end of 2026 and 20,000 by the end of 2027. The growth will be funded by $5 billion in borrowed money raised in October from private lenders without Alphabet guarantees. Waymo said it raised the money to grow in the US and abroad, and declined to comment on specific plans. Waymo currently operates in 15 US cities, and Semafor also mentions an expansion into Toronto.

Sources 1

semafor.com

Brain-computer interfaces are being fitted into ordinary clinics step by step: implantation will be planned and checked on standard scanners.

Paradromics and Siemens Healthineers to build a medical imaging system for brain-computer interface implantation

Brain-computer interfaces are being fitted into ordinary clinics step by step: implantation will be planned and checked on standard scanners.

The companies announced a collaboration on a platform meant to cover the whole implantation path: patient selection, finding the target brain area, planning and navigation during surgery, localization after implantation, and assessment of results. The workflow will combine structural and functional MRI, vascular imaging and CT, and is expected to be deployed in clinics on Siemens Healthineers equipment. Speech and motor functions come first, hand movement and restoration of speech after stroke later. Paradromics CEO Matt Angle: "A brain-computer interface's performance begins before the implant ever reaches the brain."

Sources 1

biospace.com

A large contractor is moving thousands of engineers to building SI agents, but so far the results have been named only by the company itself.

Cognizant assigns 18,000 employees to build SI agents on Google's Gemini platform

A large contractor is moving thousands of engineers to building SI agents, but so far the results have been named only by the company itself.

The new Gemini Frontier Capability Group will work on modernizing legacy systems, developer tools, business process automation and cybersecurity. More than 55,000 Cognizant employees have already been trained in working with Gemini. Financial terms of the partnership were not disclosed; CEO Ravi Kumar said the group will let the best engineers focus on bringing frontier SI into production for clients. The company cites the example of an unnamed retailer: a 40% increase in developer productivity and a 35% cost reduction.

Sources 1

thenextweb.com

The US markets regulator is preparing to discuss what data should look like when SI uses it.

SEC to hold a November 9 roundtable on machine-understandable data and "superintelligence" in financial markets

The US markets regulator is preparing to discuss what data should look like when SI uses it.

The roundtable will take place on November 9 from 9:00 a.m. to 3:30 p.m. Eastern time at SEC headquarters in Washington and will be streamed on the commission's website. The topics are machine-understandable data, "superintelligence" (in the SEC text, Super Intelligence, SI) and the related risks and costs. SEC chief economist Joshua White: "In the age of SI, high-quality information is critical." The announcement contains no panel program.

Sources 1

sec.gov

Good caching and the right model choice can make SI code review three times cheaper.

Cloudflare cut the cost of an SI code review run from $2.43 to $0.77, mostly through caching and model choice

Good caching and the right model choice can make SI code review three times cheaper.

According to Cloudflare, the cost of one review fell 68%, and median time went from 4 min 9 s to 1 min 32 s. The cache hit rate on the first request rose from 14-26% to 91-99.5% after the system prompts were restructured; the rest came from routing: Claude Opus 5.5 coordinates, Claude Sonnet 5.5 reviews, and the effort level is matched to the review's complexity. The number of important findings per review, the company says, did not fall (0.28 to 0.30), but the samples are small (35-66 findings per group), and some reviewers at medium effort lost findings.

Sources 1

blog.cloudflare.com

There is no decision yet: the commission does not plan to change the rules on SI-voiced calls, and politicians argue about the risk of deceiving voters.

Schumer urges the FCC to reject a petition that would allow SI-voiced robocalls before the midterm elections

There is no decision yet: the commission does not plan to change the rules on SI-voiced calls, and politicians argue about the risk of deceiving voters.

Senator Chuck Schumer demanded that the Federal Communications Commission reject a petition from the group Club for Growth, filed on August 31; the public comment period ran from September 4 to October 19. Schumer warned that campaigns could use opponents' voices for messages that are "at best, disingenuous and, at worst, a blatant lie." An FCC spokesperson told Al Jazeera that no one at the commission is interested in moving to an order on this or other petitions and there are no plans to do so. Club for Growth's position was not found in the sources checked.

Sources 1

aljazeera.com

To power SI data centers, tech giants are now investing in their own nuclear plants.

Amazon to fund a 960 MW nuclear plant of 12 small reactors in Indiana for its data centers

To power SI data centers, tech giants are now investing in their own nuclear plants.

The plant of 12 X-energy Xe-100 reactors will be built on the site of a demolished coal plant south of Terre Haute; the operator is American Electric Power. Amazon will support development with funding and a long-term power purchase agreement. This is Amazon's second commercial agreement with X-energy, three times larger than the first; the parties' goal is 5 GW of new nuclear capacity in the US by 2039. The plant will be connected to the public grid and supply Amazon's data centers. The timeline depends on completion of X-energy's project in Seadrift, Texas, for which the company is awaiting final regulatory permission to build.

Sources 1

techcrunch.com

For document processing, a model appears that answers faster and cheaper than language models, but the figures so far come only from the company itself.

Cloudflare introduces Clairic, a model that pulls field values out of text and images in a single pass without generating text

For document processing, a model appears that answers faster and cheaper than language models, but the figures so far come only from the company itself.

Clairic takes a schema of typed fields (text, number, date, duration, list, estimate, forecast) and returns values with a confidence score, and if the data is missing it returns an empty value rather than an invented answer. The model is available in preview in Workers AI under the name @cf/cloudflare/clairic; Cloudflare promises open weights later. By the company's own measurements, a claims-processing pipeline with Clairic took 19.3 s against 61.7 s for a pipeline of language models alone and cost 49% less ($6,032 against $11,740 per million claims); in a 21-day forecast the average error was 3.3%, with 17 of 21 actual values falling within the 80% interval.

Sources 3

blog.cloudflare.comthe-decoder.comblog.cloudflare.com

Video on demand is generated in seconds and costs pennies, so trying out an idea will be easier for anyone.

fal and Black Forest Labs release FLUX 3 Max: a ten-second video is generated in about 12 seconds

Video on demand is generated in seconds and costs pennies, so trying out an idea will be easier for anyone.

According to VentureBeat, the model was refined by the fal Research team together with Black Forest Labs on the basis of FLUX 3. By fal's measurements, a 10-second clip is generated in 12.45 s against 277.80 s for Seedance 2.5 - about 22 times faster. On the Artificial Analysis image-to-video leaderboard as of October 19, FLUX 3 Max scores 1,138 Elo and shares first place with Dreamina Seedance 2.5, but the confidence intervals overlap heavily, and by point estimate the place is second, with a possible range from first to fifth. Text-to-video, image-to-video and keyframe generation are available through the API, with clips of 5 to 20 seconds at 720p; with a 75% discount until November 7, 10 seconds cost $0.425 (regular price $1.70).

Sources 1

venturebeat.com

An algorithm helped find a molecule against a dangerous bacterium, but a drug is a long way off: experiments in living organisms are needed.

Peptide CIT8, refined by a machine-learning algorithm, killed the antibiotic-resistant bacterium Acinetobacter baumannii in the lab within 15 minutes

An algorithm helped find a molecule against a dangerous bacterium, but a drug is a long way off: experiments in living organisms are needed.

In a Frontiers in Microbiology paper, researchers at Houston Methodist Hospital and partners described a 13-residue peptide, CIT8, optimized by a machine-learning algorithm. It stayed active (minimum inhibitory concentration of 4 ug/mL) against 15 of 16 tested A. baumannii strains, including clinical isolates resistant to carbapenems; at 20 ug/mL (5 times the minimum inhibitory concentration) a population of about 5x10^7 CFU/mL fell below the detection limit within 15 minutes. Together with meropenem the peptide showed synergy. The work is preclinical and the experiments were done in vitro: the authors write that proteolytic stability, pharmacokinetics and the risk of resistance have not been studied, and that activity dropped in the presence of divalent cations and human serum; experiments in living organisms are needed.

Sources 2

frontiersin.orgixbt.com

Disputes over whether models may be trained on news without a licence have reached another country, and there is no answer yet.

Canadian publishers Postmedia, Torstar, The Globe and Mail and The Canadian Press sue Anthropic over training Claude on their texts

Disputes over whether models may be trained on news without a licence have reached another country, and there is no answer yet.

According to MobileSyrup, the suit was filed on October 6 in the Ontario Superior Court. The plaintiffs allege that Anthropic copied their materials without a licence to train Claude, circumvented technical protection measures and violated the websites' terms of use; the statement of claim says the company "capitalized on the commercial success of its models" without holding a licence. The same publishers, per the same source, have been suing OpenAI since 2024 - the case is unresolved. Anthropic's position was not found in the sources checked.

Sources 1

mobilesyrup.com

By IFP's calculations, China's compute for SI will nearly double by 2027, with a large share coming from US chips and remote access to them.

IFP estimates China's computing resources at 4.6 million H100-equivalents by the end of 2026 and 8.8 million by the end of 2027; US chips will make up about 65%

By IFP's calculations, China's compute for SI will nearly double by 2027, with a large share coming from US chips and remote access to them.

The Washington think tank Institute for Progress (IFP), in a report on the US–China gap in SI, puts the lag of the best Chinese models behind American ones at about 6.5 months. By its calculations, China's access to compute will grow from about 4.6 million H100-equivalents at the end of 2026 to 8.8 million at the end of 2027 — from domestic production, legally imported and smuggled US chips, and remote access to them in data centers abroad. US chips will make up about 65% of that stock; of the resources not yet in China, about half will come from remote access. IFP also estimates that "distillation" of US models lets Chinese companies recapture about half of the lag. The authors propose measures within existing authorities — export controls on remote access to training clusters abroad, stricter due diligence against smuggling, and an antitrust "safe harbor" for joint efforts against distillation — and believe they could roughly double the US lead by the end of 2027. These are the estimates and recommendations of a single think tank; the report does not give the position of the Chinese side or of chip makers.

Sources 1

ifp.org

New digital command systems still fit poorly with old communications, and the US Army admits it.

US Army ran into trouble at Hawaii tests linking Anduril's Lattice command system with older communications

New digital command systems still fit poorly with old communications, and the US Army admits it.

At the Lightning Surge 5 exercise in September in Hawaii, the 25th Infantry Division connected Anduril's Lattice platform, part of the NGC2 command architecture, to existing communications equipment. Brig. Gen. Shane Taylor told Breaking Defense: "It wasn't seamless, and we learned a lot." Lockheed Martin vice president Chandra Marshall, whose company is the main integrator for this division, noted that passing data through Lattice is not yet at the level it should be. The jungle, lighter vehicles instead of Strykers and differences in radio configurations from the 4th Division in Colorado were also named as complications. The Army intends to roll out a baseline version in other I Corps divisions this fiscal year, but a continuing resolution could slow the pace.

Sources 1

breakingdefense.com

If the method is confirmed in clinics, fetal MRI will be blurred less often by movement.

MIT: E(3)-Pose, a neural network that uses the symmetry of the fetal head, helps produce clear MRI scans despite fetal movement

If the method is confirmed in clinics, fetal MRI will be blurred less often by movement.

In fetal MRI, head movement means slices are captured at different angles, and subtle brain anomalies can be missed. The E(3)-Pose model, presented at the ECCV conference, tracks head position and builds symmetry into the network's architecture itself rather than learning it from data alone. The lead author is MIT PhD student Ramya Muthukrishnan. The model was trained mostly on third-trimester fetuses; the authors want to extend it to more active second-trimester fetuses. MIT's article gives no quantitative metrics.

Sources 1

news.mit.edu

It is getting easier for SI agents to find what they need on the internet, but the leadership is so far claimed by the company itself.

Exa releases Exa 3.0, a search API for SI agents; by the company's measurements, it ranks first in all eight tests

It is getting easier for SI agents to find what they need on the internet, but the leadership is so far claimed by the company itself.

According to Exa's blog, version 3.0 scores 91.8% on BrowseComp and 88.2% on DSQA and, as the company claims, beats Perplexity, Brave and Parallel on eight metrics. The update includes a larger embedding model, a Dynamic Highlights mode (excerpts that, according to the company, use 95% fewer tokens than the full page text) and an expanded index - 1.27 billion ecommerce listings and over 211 million local businesses. Exa Lite costs $4 per 1,000 requests, Exa Auto $7.

Sources 1

exa.ai
Previous day
Follow on Telegram