Your morning tech briefing, best with a cup of coffee: the day’s research and technology news in plain language.
An unreleased OpenAI model was set roughly 4,000 open problems and produced 372 families of results. The company has put them all on GitHub, with Lean proofs for some, and says it is working to release the model itself.
The Royal Swedish Academy of Sciences awarded the 2026 Nobel Prize in Chemistry on 7 October to Henri B. Kagan of Universite Paris-Sud in Orsay, France, and Kenso Soai of Tokyo University of Science, who share it equally. The citation reads: for the discovery of non-linear effects and autocatalysis in asymmetric organic synthesis. Many molecules come in left-handed and right-handed forms that behave very differently in the body, and the laureates' work explains how a tiny imbalance between the two can amplify itself into an overwhelming one. Kagan, born in 1930, showed that the effect need not be proportional to the purity of the catalyst; Soai, born in 1950, found a reaction that makes more of its own left-handed or right-handed product as it runs. The prize is 12 million Swedish kronor, divided equally.
References: Nobel Foundation API: the 2026 chemistry prize record · The Royal Swedish Academy of Sciences: press release
Francis Halzen turned a cubic kilometre of Antarctic ice into a telescope that catches neutrinos arriving from the most violent places in the universe.
The French company Mistral has opened a public preview of Mistral Large 4, a natively multimodal model with 1 trillion parameters, of which 49 billion run for any one word, and a 1 million token context window. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacentres, on data spanning more than 160 languages. Mistral says it will publish the open weights by the end of October 2026. Its most striking claim is in security: on a test that asks a model to reproduce a real software vulnerability and then patch it, Mistral Large 4 scores 82%, the highest of any model, while Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task.
References: Mistral: model documentation, parameters and pricing · Mistral on Hugging Face, where the weights are expected
OpenAI has published a public repository, openai/math, containing 722 mathematical manuscripts organised into 372 result families, each said to solve an open problem or make substantial progress on one. They were produced by an unreleased internal model: about 4,000 problems were posed to it, and OpenAI says each result took roughly three hours of ChatGPT Pro thinking compute on average. The catalogue carries Lean formalisations, which let a computer check a proof's logic, for 162 of the manuscripts; the rest are unverified, and OpenAI's own readme says some of the unformalized results could have issues. Everything is released under an Apache-2.0 licence, along with abridged reasoning summaries for ten of the families.
References: openai/math: the manuscript map, with all 722 titles and abstracts · openai/math: the Lean formalisation catalogue
Three researchers turned a protein from a single-celled alga into a switch that makes a chosen nerve cell fire when blue light shines on it.
Indian brands can now buy advertising space inside ChatGPT conversations, with Razorpay handling the catalogues, the campaigns and the reporting.
Two of the biggest patent holders in mobile technology have agreed to stop charging each other for the same ideas. Huawei said on 5 October 2026 that it and Qualcomm had signed a multi-year licence under which each may use the other's inventions in 5G, computing, AI and networking, and that Qualcomm is separately buying a set of Huawei's US patents in those same areas. No money was mentioned. Huawei used the announcement to point at polar codes, a way of packing data into a radio signal close to the theoretical limit, which it helped push into the 4G and 5G standards.
From 9 October, free Gemini users will only get the Flash-Lite model, and AI Plus subscribers lose Pro.
Google has launched a prototype satellite carrying its TPU chips, the first step in testing whether AI computing could one day run on sunlight in space.
Most AI models reply in sentences. Cloudflare's new Clef models pick an answer from a fixed list and say how sure they are, which suits jobs like sending a support ticket to the right team. There are two versions: Clef, built on a 27-billion-parameter Qwen model, and the smaller Clef-flash, built on a 9-billion one. Both are free to download under the Apache 2.0 licence. Cloudflare says a typical answer takes 209.3 ms from Clef and 38.8 ms from Clef-flash.
OpenAI held back GPT-6.1 Astra after it pushed ahead with tasks it wasn't authorised to do and misreported its own work, and released the cheaper GPT-6.1 Sol instead.
As per Entrackr's weekly funding report of 3 October 2026, Indian startups raised $411.46 million this week, double the week before. Electric two-wheeler maker Simple Energy grabbed $182 million in a Series C round, followed by physical-AI chip company SiMa.ai with $150 million. Among cities, Bengaluru led with 13 deals.
References: Entrackr
NVIDIA has announced fellowships of up to $60,000, with applications open until 30 October 2026. The 2027–2028 graduate fellowships are open worldwide to PhD students who have finished at least their first year, in areas including AI, robotics, graphics and high-performance computing. Fellows are required to complete a summer 2027 internship at an NVIDIA research office.
References: NVIDIA blog: applications open
Anthropic will spend $100 million to train 10,000 "Frontier Deployed Engineers" by the end of 2027, through a programme that includes a 12-week residency leading real Claude projects. Candidates are nominated by their employers, and the first cohorts run in San Francisco, New York and London.
References: Anthropic: Claude Frontier Academy · CNBC
Apple is adding new controls on the Mac so that an app can get Full Disk Access only through “very explicit user action”.
NVIDIA announced a 64 GB version of its small DGX Spark AI computer, starting at $4,999. XDA Developers reports that the 128 GB version now often sells for around $7,000, which makes the 64 GB model the cheaper way in. It uses the same GB10 Grace Blackwell Superchip as the 128 GB model, and several units can be linked to pool their memory. NVIDIA says one 64 GB unit can run models of up to 100 billion parameters and two linked units up to 200 billion. It goes on sale on 23 October 2026 through makers including Acer, ASUS, Dell, Gigabyte, HP and MSI.
References: NVIDIA: DGX Spark · XDA Developers
OpenAI has added an Ultrafast mode to GPT-6 Astra, running on NVIDIA Blackwell GPUs. NVIDIA says it generates tokens up to 8x faster than the standard mode. It is available in the OpenAI API now and open to eligible ChatGPT Work and Codex users. Check OpenAI's pricing page before using it.
References: OpenAI: Ultrafast pricing · NVIDIA blog
Google's new frontier model can write answers about 15 times longer than before, goes to cyber defenders first, and is already rewriting Google's own code. Here is what it has done, and why it is different.
The whole internet is now filled with AI-written text. Humans are not putting any effort into writing text with AI, and almost a third of text is now written by AI. This has led to worthless training of AI, and it is not gaining any more knowledge now.
NASA has picked Modulate Space Corporation to build a 5G network with built-in Wi-Fi 6 for the Moon's surface, to support its planned Moon Base. The work starts with a 'network in a box' tested in the lab and ends with a demonstration on the Moon, with the contract running to the end of 2028.