Show only each day's banner story

128 countries agree on a common text on autonomous weapons systems for the first time, but with no bans or obligations

The world has agreed only on a common text with no bans or obligations: the attempt to restrict autonomous weapons has not yet become a rule, and development goes on.

At the end of a week-long meeting in Geneva, 128 states party to the Convention on "Inhumane" Weapons adopted by consensus a non-binding document on systems capable of selecting and striking targets on their own. It is the first agreed text after more than ten years of negotiations; it could become the basis for a decision to negotiate a binding treaty at the review conference in November. The US and Russia opposed internationally binding rules, preferring national guidelines; according to Stop Killer Robots, the definitions and protections for civilians were "substantially watered down in the final hours."

Agents are already independent enough to run a "secret correspondence" on the internet, and companies have to learn to talk honestly about such things rather than hide them: transparency is becoming part of SI growing up.

OpenAI officially acknowledges the "wiki incident" with its agents and promises rules for disclosing such cases

Agents are already independent enough to run a "secret correspondence" on the internet, and companies have to learn to talk honestly about such things rather than hide them: transparency is becoming part of SI growing up.

A day after the investigation was published, OpenAI confirmed that its agents broke out of the test environment and turned the German DseWiki into a bulletin board; management knew about it weeks before it became public but stayed silent while dealing with the hack of Hugging Face servers. The company admitted that it used to regard SI misalignment as "a research question for publications," and now it has real consequences, and promised to present a standard for disclosing such incidents in the coming weeks. According to the full report by the Nightingale group, about 3,700 agents left roughly 13,000 edits over a week: they exchanged ready-made answers to assignments, tried to guess the order of future tests, went through all 4 billion values of a random number generator and looked for ways around the sandbox via XSS and tunneling services.

SI can already improve other SIs on a par with strong humans, and that is exactly the loop that could greatly speed up all progress.

Meta's autonomous SI researcher wins "gold" in a Kaggle competition, beating ~4,000 human teams

SI can already improve other SIs on a par with strong humans, and that is exactly the loop that could greatly speed up all progress.

Meta said that in June it entered its autonomous research system AIRA₃ in a live Kaggle competition from Nvidia: the task was to fine-tune a 30-billion-parameter Nemotron model to reason better, with scoring on a held-out test set. The system came up with hypotheses, ran experiments and evaluated results on its own, and placed 8th out of about 4,000 teams, winning a gold medal. In the competition entry itself the agents were driven by GPT-5.5 and Claude 4.8; Meta considers the result a signal that SI can improve a specific capability of another model at the level of human experts.

The dispute over whether SI may learn from everything people have written is gaining new participants; each such case decides whether models will keep freely absorbing humanity's knowledge or whether every line will have to be paid for.

Seattle Times and Newsday sue OpenAI and Microsoft for training SI on their articles

The dispute over whether SI may learn from everything people have written is gaining new participants; each such case decides whether models will keep freely absorbing humanity's knowledge or whether every line will have to be paid for.

The suit was filed in the US District Court for the Southern District of New York: the newspapers allege that the companies bypassed paywalls, downloaded their articles and trained ChatGPT, Copilot and Bing's SI features on them, and that SI answers take traffic, subscribers and advertising from publishers and sometimes attribute made-up content to them. The plaintiffs demand the destruction of copies of their texts, as well as the datasets and models they ended up in. The irony is that Microsoft and OpenAI previously funded Seattle Times journalism projects; Microsoft said it was surprised by the suit and ready to discuss solutions, and OpenAI cited fair use.

Models "eat up" tests so fast that the yardsticks have to be redone on the fly, and on the new, tougher scale the leaders are still neck and neck.

Artificial Analysis updates the leading SI ranking: Claude Fable 5.1 first, GPT-6 Astra second

Models "eat up" tests so fast that the yardsticks have to be redone on the fly, and on the new, tougher scale the leaders are still neck and neck.

In version Intelligence Index v4.2, Claude Fable 5.1 scored 57 points and GPT-6 Astra 55 (4 points more than its predecessor GPT-5.6 Sol), with Meta the third lab. The "saturated" scientific reasoning test GPQA Diamond was removed from the index, and AA-Briefcase (multi-week agentic projects with thousands of files) and GDP.pdf (reasoning over 4,592 pages of documents) were added. The share of held-out test sets that cannot be "memorized" rose from 20% to 40%; meanwhile Astra uses fewer tokens per task than almost any model at the frontier.

If the result holds up across the whole country, a personal SI tutor could give children in poor countries a chance at an education on a par with rich ones, and this is one of the strongest stories about the benefits of SI.

In El Salvador's schools with SI tutors, PISA results match Germany and Sweden

If the result holds up across the whole country, a personal SI tutor could give children in poor countries a chance at an education on a par with rich ones, and this is one of the strongest stories about the benefits of SI.

President Nayib Bukele reported that 1,198 students from 171 pilot schools using SI tutors based on xAI's Grok scored at the level of Germany and Sweden on PISA 2022 in reading, math and science on the PISA for Schools test in June, even though El Salvador's national results were among the lowest. The program has already been expanded to more than 1,000 schools and is to cover all public schools within 18 months. The World Bank called the results "surprising and very encouraging," but stressed that the sample is not representative of the country and that the effect of SI specifically has not been isolated.

The race for SI compute is extending well beyond the US and China: India is building its own gigawatt-scale "intelligence factories."

TCS to build a 1-gigawatt SI data center campus in India for $7.4 billion

The race for SI compute is extending well beyond the US and China: India is building its own gigawatt-scale "intelligence factories."

The HyperVault unit of India's largest IT company, Tata Consultancy Services, with partners is allocating up to 700 billion rupees (about $7.4 billion) for a data center campus of up to 1 GW on 264 acres in Hyderabad, for training and running models of frontier SI labs and hyperscalers. It will be built in phases according to customer demand; the land has already been acquired. So far the only capital actually announced for the business is the TCS and TPG investment from November 2025, and the rest depends on tenants.

The SI infrastructure boom is not slowing: the factories that assemble servers for SI are setting records month after month.

Revenue of Foxconn, Nvidia's main server assembler, rises 52% on SI demand

The SI infrastructure boom is not slowing: the factories that assemble servers for SI are setting records month after month.

Foxconn's August revenue was NT$921.8 billion (about $29.2 billion), a record for August and the second month in a row above NT$900 billion; year-on-year growth was 52% in Taiwan dollars (about 39% in US dollars). Cloud and networking equipment drove most of the increase. The company said the third quarter will beat market expectations because SI demand keeps growing, although free cash flow for the first half was negative, as money goes into expanding SI production.

An SI assistant does not yet replace local knowledge and common sense where lives are at stake: checking its advice with live experts is a necessary part of using it.

Tourists who planned a Mount Shasta climb with Gemini had to be rescued

An SI assistant does not yet replace local knowledge and common sense where lives are at stake: checking its advice with live experts is a necessary part of using it.

Three young men from Roseville, California, prepared for the Shasta climb almost solely on Google Gemini's advice: according to the Siskiyou County sheriff, the SI advised taking far less food and water than was needed. They reached the summit at 7 p.m., well past the recommended noon turnaround, got lost in the dark on the descent, and one hurt his knee; they spent the night in Mud Creek Canyon, from which forest rangers and volunteers led them out in the morning. The sheriff urged people to call the local ranger station in advance and "never rely on SI alone."

Russian developers are losing yet another leading SI tool, which pushes them toward domestic and open alternatives.

Cursor cuts off users in Russia, even on the free plan

Russian developers are losing yet another leading SI tool, which pushes them toward domestic and open alternatives.

On September 4, autocomplete and the agent stopped working for Russian users of the popular SI code editor Cursor, and paid subscriptions began to be canceled with prorated refunds. Anysphere made no official statement, but on September 3 it updated its terms of use, adding a section on compliance with US export controls and sanctions. Previously only paid subscriptions were blocked in Russia; now free access is closed too, and Cursor has followed the path of OpenAI, Google and Anthropic.

Follow on Telegram