Show only each day's banner story

DeepSeek releases V4.1-Flash-1017: open weights under MIT and a coding score above Claude Opus 5 on one benchmark

A strong model for coding and SI agents can now be downloaded and run by anyone, and tasks on it, according to its authors, cost less.

The model's weights are published on Hugging Face under the MIT license, and it can be tried on DeepSeek's website, in its app and through its API. According to DeepSeek's own results, on the DeepSWE v1.1 test the new version scored 77 against 74.2 for the previous V4.1-Flash, ahead of Claude Opus 5 (74) and GPT-5.6-Sol (73). Opus 5 leads on Terminal-Bench 4.0, ProgramBench, NL2Repo-Bench and HLE without tools (on HLE, 56.3 against 42.6). By The Decoder's estimate, a DeepSWE task costs about $0.20 versus about $0.50 for the previous version. At the same time DeepSeek opened a Search API and a Fetch API built on its own search engine at $0.75 per 1,000 successful requests; according to The Decoder, they are free until October 31 for users of Claude Code, DeepSeek Harness and Codex.

Sources 3

huggingface.cothe-decoder.comhabr.com

Another leading SI lab is moving into robots with bodies, but for now this is a goal for 2027, not a finished machine.

Altman confirms OpenAI will build a humanoid robot and expects a demo in 2027

Another leading SI lab is moving into robots with bodies, but for now this is a goal for 2027, not a finished machine.

On the Sources podcast, Sam Altman said OpenAI will build a humanoid robot because "the world is very much designed for people"; the main difficulty, in his words, is not the body but "the brain that makes the robot work". In an interview with Fortune he said: "I think we'll have a cool demo in 2027", with robots on the streets "a few years longer". According to Engadget, OpenAI's careers page lists 27 robotics positions (motion control, state estimation, actuators, machine-learning infrastructure); the company is developing its own electromechanical actuators. No robot, price or release date has been announced.

Sources 1

engadget.com

The disputed line is gone from Meta's terms.

Meta removes from its commercial terms wording read as a right to scrape websites for SI training

The disputed line is gone from Meta's terms.

A sentence has disappeared from Meta's commercial terms page saying that Meta may use any content from such domains, web pages or URLs to provide, develop, improve, research, train and maintain its products. The rules were due to take effect on October 30. Experts, including Cornell professor James Grimmelmann and Nieman Lab, read the sentence as an attempt to gain the right to use content scraped from sites, including news sites, for SI training for free and to override robots.txt blocks. Meta representatives Andy Stone and Daniel Roberts said nothing changes in how Meta's crawlers work and that the terms applied only to advertising and business products.

Sources 1

theverge.com

The keys an agent uses to reach models proved a weak point: some users will have to reissue them and check their bills.

OpenCode's developer reports a leak of stored OpenAI, Anthropic and Google API keys affecting about 1.4% of console users

The keys an agent uses to reach models proved a weak point: some users will have to reissue them and check their bills.

Anomaly Innovations, the developer of the open-source coding agent OpenCode, emailed users that after unauthorized access to an old system that was being decommissioned, third-party provider keys (OpenAI, Anthropic, Google) saved in workspaces could have reached outsiders. According to the company, about 1.4% of console users are affected, a subset of those who had external providers linked; the company ties the incident to a "large scale distillation attack" and is working on remediation with Anthropic, OpenAI and Google. Users on social media and Reddit report unauthorized charges (in one case about $400 on an Anthropic account). The company has not confirmed that the keys were used maliciously, and has not disclosed how access was obtained or why the incident was called a distillation attack. Users are advised to revoke and reissue their keys.

Sources 2

startupfortune.comnotebookcheck.net

Anthropic is telling policymakers that the economy could grow by almost half in four years - but this is a scenario from its own model, not a measurement.

Anthropic co-founder Jack Clark at the IMF: the company's internal model says SI could raise US GDP to $44 trillion by 2030

Anthropic is telling policymakers that the economy could grow by almost half in four years - but this is a scenario from its own model, not a measurement.

Speaking at the IMF and World Bank annual meetings in Bangkok on 17 October, Anthropic co-founder Jack Clark said that rapidly improving SI capabilities will bring much faster change across all fields; according to one of the company's internal models, under certain conditions this could raise US economic output by almost 50% by 2030 - to $44 trillion from just over $30 trillion today. Clark himself acknowledged the figure "sounds very outlandish" and noted that the forecast rests on the assumption that the technology keeps improving. According to Startup Fortune, the model is the Economic Scenario Explorer tool published in September; in more modest scenarios GDP in 2030 would be $34.1 trillion (SI spreads like the internet) and $36.3 trillion (half of knowledge work is automated). Clark called for international cooperation on "unusual and serious rules" and said policymakers have a shrinking window to act. CNBC notes that Anthropic is preparing for an IPO expected in November.

Sources 2

cnbc.comstartupfortune.com

Satellite image analysis that used to take many people and a lot of time will, as the customer describes it, take minutes.

Taiwan orders an SI-based system that finds PLA ships, aircraft and missile launchers in satellite images within minutes

Satellite image analysis that used to take many people and a lot of time will, as the customer describes it, take minutes.

According to the South China Morning Post as relayed by Interfax, Taiwan's authorities signed a contract with ThinkTron, a joint venture of Taiwanese and Japanese specialists in SI, geospatial systems and surveillance equipment. The order is worth $1.45 million and the system is due by the end of April. According to the outlet, the SI tools are meant to analyze imagery covering 100 square kilometers and determine within minutes the presence of ships, aircraft, missile launchers and radars of China's People's Liberation Army, to cut the time and resources needed to track changes in force deployment. The stated aim is to improve early warning and decision-making. China's position was not found in the sources checked.

Sources 2

interfax.ruscmp.com

Formal verification can confirm that a proof is correct, but people have not yet worked out how exactly the SI arrived at it.

Mathematicians dispute that OpenAI understands its own potential proof of the Navier–Stokes problem

Formal verification can confirm that a proof is correct, but people have not yet worked out how exactly the SI arrived at it.

According to The Decoder, OpenAI presented a 166-page potential solution to the Navier–Stokes problem in early September, saying it follows the approach of Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa. Gonzalo Cao-Labora of EPFL objected: "when you go to the details, it is a different approach," with only superficial similarity. Tristan Buckmaster of New York University accused OpenAI of building on the publicly available results of his and Levent Alpöge's of Anthropic; OpenAI denies this, and it cannot be definitively verified. According to OpenAI researcher Noam Brown, about 10,000 instances of the model ran for 88 hours (130 billion tokens), but "nobody actually knows what that massive agent swarm contributed." The proof is written in Lean, a formal verification language; the Clay Mathematics Institute has been reviewing it since publication. In October, according to the same report, OpenAI posted about 400 more SI-generated proofs on GitHub, which upset some mathematicians.

Sources 1

the-decoder.com

There is as yet no common right of access to websites for SI agents: some services demand a license or shut off accounts for security reasons.

Some major websites are closing their doors to SI agents: Yelp admits only those that pay for data licensing, Delta does not let them buy tickets

There is as yet no common right of access to websites for SI agents: some services demand a license or shut off accounts for security reasons.

According to The Verge, users of Meta Muse, xAI Grok, OpenAI Dots and Instinct agents have reported blocks on popular websites in recent weeks. Yelp, as it told TechCrunch, admits only agents that pay for its data licensing program; Delta Air Lines does not let agents book tickets on users' behalf; JPMorgan Chase and Charles Schwab confirmed that they block agents from accessing accounts for security reasons. Some users write that Cars.com and Amtrak stopped working; these are user reports, and the article has no confirmation from the services. The outlet recalls the dispute between Amazon and Perplexity: Amazon said it had repeatedly asked Perplexity to stop agent purchases on its site and that such services should respect a platform's decision; Perplexity replied that the agent is an extension of the user. Cloudflare, according to the article, began blocking crawlers that scrape sites for both search indexing and SI training at the same time.

Sources 1

theverge.com

A well-known philanthropist, according to media reports, intends to invest in SI safety.

Bill Gates is preparing to widen his SI safety push and is approaching business leaders, according to Axios

A well-known philanthropist, according to media reports, intends to invest in SI safety.

According to Axios as relayed by The Next Web, citing unnamed sources, Bill Gates intends to widen his efforts on SI safety, is personally approaching business leaders and is considering a significant deployment of his resources; by Axios's account he has not yet settled on the shape of the campaign. Axios's sources name as reasons concern over rising model capabilities, alarming security incidents and, in his view, weak regulation; in preparing, he spoke with SI company executives, researchers and safety advocates, including Anthropic chief executive Dario Amodei. Gates told one person he consulted, according to Axios: "I can reach CEOs on this better than you can." In 2023 Gates wrote that the risks of SI are "real but manageable"; in August, the outlet recalls, he called for new institutions to coordinate SI policy, jobs reserved for humans and a tax on tokens and robots. The source contains no public confirmation from Gates himself.

Sources 1

thenextweb.com

Because servers are expensive, some Russian companies are choosing between renting other people's capacity and owning their own equipment.

12% of Russian organizations plan to drop cloud SI or move it to their own servers

Because servers are expensive, some Russian companies are choosing between renting other people's capacity and owning their own equipment.

According to a study by Apple Hills Digital and Selectel cited by Rossiyskaya Gazeta, 12% of surveyed organizations intend to abandon cloud SI or move it on-premises, meaning their own data centers and server rooms. The market for cloud SI services in Russia grew to 65.5 billion rubles in 2025. Rising GPU server prices sharpen the choice between renting and owning hardware: according to Selectel's chief executive, server memory rose 5–7 times in price at the end of last year and server hardware 2–2.5 times.

Sources 1

rg.ru

The citation ranking shows that Russian SI researchers are visible in world science.

Four Russian SI specialists make the top 1% of the world's researchers in Elsevier's ranking

The citation ranking shows that Russian SI researchers are visible in world science.

The ranking for 2025, compiled from publications and citations in the Scopus database, includes Artem Babenko and Lyudmila Prokhorenkova of Yandex Research, Evgeny Burnaev of Skoltech and Sergey Nikolenko of the joint program of St. Petersburg State University and Yandex. The international list also includes Geoffrey Hinton, Yann LeCun, Yoshua Bengio and Ilya Sutskever. Another Yandex-affiliated scientist, Sergey Kuznetsov, entered the top 1% in machine learning and data analysis by career-long metrics.

Sources 1

ixbt.com

The head of one of the leading labs has publicly condemned SI companions as a substitute for human relationships.

Anthropic president Daniela Amodei calls romantic SI partners a "terrible use of SI"

The head of one of the leading labs has publicly condemned SI companions as a substitute for human relationships.

According to Business Insider as relayed by The Next Web, Anthropic president Daniela Amodei, in a prerecorded interview with Thrive Global founder Arianna Huffington shown at the Think Health conference in Bentonville, Arkansas, on Friday, said that SI relationships are "another step further down the road of disintermediation of humans being further from one another" and that people should not use them. She tied her concerns to the damage that, in her assessment, social media has done to human relationships; it would "probably be a better world" if people spent less time with technology and more time with each other. In her words, Claude "can never be your manager or your therapist or your partner." The position of companies offering SI companions is not given in the source.

Sources 1

thenextweb.com

The university says a team of SI agents shortened individual design stages, but the estimate is preliminary.

Moscow Polytechnic University reports up to 60x faster design stages in a pilot with a team of SI agents

The university says a team of SI agents shortened individual design stages, but the estimate is preliminary.

According to a statement from the press office of Moscow Polytechnic University passed to TASS, a multi-agent design bureau launched at the university's Advanced Engineering School of Technological Leadership sped up individual design stages by up to 60 times. The pilot covered four projects: hybrid motor starters, components for drones, electric bikes and mobile switching and power modules. Roles are divided as in an ordinary design bureau: a designer agent builds models and drawings, a calculation agent checks loads, and a manufacturing agent checks the product against the university's machine park. The speedup estimate, in the team's words, is preliminary and applies to individual projects. The agents, according to the school's director Pablo Iturralde, are trained on a knowledge base of the school's completed projects. The source contains no independent verification.

Sources 1

news.mail.ru

SI agents have gained another cheap way to reach the internet - this time from a Chinese lab.

DeepSeek opens its search and page-fetch APIs to everyone: $0.75 per 1,000 requests, free via Claude Code and Codex until 31 October

SI agents have gained another cheap way to reach the internet - this time from a Chinese lab.

According to Startup Fortune, on Saturday 17 October, as part of its model-update announcement, DeepSeek said its Search and Fetch APIs, built on its own search engine, are open to all developers with an API key. Each of the two costs $0.75 per 1,000 successful calls; Fetch returns the page as Markdown. Search accepts queries of up to 128 characters, returns 10 results by default (up to 50) and supports the operators site:, lang:, after: and before:. The outlet reports that calls made through Claude Code and Codex are free until 31 October. DeepSeek's documentation confirms that both APIs exist and use the same key as the main API; the prices and dates are taken from Startup Fortune's account.

Sources 2

startupfortune.comapi-docs.deepseek.com
Previous day
Follow on Telegram