Uncategorized

DailyAI Briefing: An AI agent operates on a mouse, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. An AI agent has injected a gene therapy into a living mouse’s brain, and the most telling moment came when it refused to bend its own rules. Aramco has closed its $2.6 billion deal for nearly half of Saudi Arabia’s national AI company, DeepSeek has shared the design for an agent that never clocks off, OpenAI is moving ChatGPT and Codex onto one shared allowance, and ChatGPT will soon be able to end chats that fixate on violence.

Let’s dig in.

An AI agent operates on a mouse

The agent chose one gene therapy from 19 options, picked the spot in the brain, worked out the coordinates from three papers and then drove the needle itself, injecting 250 nanolitres into a mouse with epilepsy. That’s the account from Gabriel Ocana-Santero, chief executive of the San Francisco biotech Exin Therapeutics, in a thread on X and an interview with WIRED. A veterinary surgeon opened the skull first, approved every irreversible step and kept a hand on a kill switch for the motors.

The model choice is the interesting part. Claude Opus 5.5 planned the work but refused to guide the needle, so the team switched to Claude Sonnet 5, which agreed. Then a glitch threw off the rig’s coordinates. The agent spotted that the position was wrong but would not touch the equipment code, because its instructions said not to. The team got a separate agent to make the fix, without telling it a live surgery was under way, which says something about how people will work around a cautious AI. The mouse is recovering, nobody knows yet whether the therapy cuts its seizures, and none of it has been peer reviewed.

Aramco closes its $2.6 billion HUMAIN deal

Aramco now owns 49% of HUMAIN, Saudi Arabia’s national AI company, with the Public Investment Fund keeping 51%, the two announced on Sunday. Aramco is paying about SAR 9.9 billion ($2.6 billion) in cash this year and handing over its own AI, computing and digital assets. The deal formalises a term sheet signed a year ago, Arab News notes.

HUMAIN, launched in May 2025, wants 3 gigawatts of AI data centre capacity by 2030 and aims to become one of the world’s leading exporters of AI. In plain terms, the world’s biggest oil company is betting part of its future on selling computing power the way it sells crude. With investors getting nervous about AI spending this week, the Gulf’s money looks steadier than Wall Street’s.

DeepSeek designs an agent that never clocks off

DeepSeek’s Harness team published a long design piece on X on Sunday describing a “super-agent” that keeps context across sessions, runs several jobs in parallel and stays online all the time. It also open-sourced a working research prototype under the MIT licence: a coding agent that watches one GitHub repository, picks up work when someone allowed asks, does it in its own session and replies, for example with a pull request.

DeepSeek is careful to call it reference code, not an official product, and says much of it was written by coding agents. The guiding rule is a good one: take as little human attention as possible and give as much value as possible. It’s pitched at the same work as Claude Code and Codex, and it comes from a lab that was reported this week to have trained on outputs from OpenAI’s GPT-6 Astra.

ChatGPT and Codex will share one allowance

From December, Plus, Pro and Business subscribers will draw on one allowance for both ChatGPT and Codex, measured in tokens, according to a newly updated OpenAI Help Center page. OpenAI says the change comes as it brings the agent features of Work into the main ChatGPT app. Bigger models and harder tasks will use more of the allowance, and quick Instant chats won’t count at all.

OpenAI says it expects most people won’t hit a reset more often than they do now, but it hasn’t given token totals, plan figures or an exact start date. That’s the catch. Heavy Codex users and heavy chatters will now be spending from the same pot, and when it runs dry the options are to wait, upgrade or buy credits. Watch the Usage page in Settings in December.

ChatGPT will end chats that fixate on violence

OpenAI made more than 200 referrals to law enforcement in the third quarter, it said in a safety update late last week, and it is now adding new limits. When a conversation shows someone fixating on violence or on seriously harming another person, ChatGPT will be able to end the chat automatically, and several such endings may get the account paused for a while. All users will also get less detail about past mass attacks, including the weapons used and how they were obtained.

Teens get tighter rules still. ChatGPT will hold back practical information about weapons and security systems from under-18s even when a question looks harmless, and OpenAI is working on alerts so parents hear if a teen’s account is paused. The changes start rolling out in the next few weeks, and the first public Community Safety Report is due in early 2027. Given how many teens already turn to AI when they feel low, the hard part will be stepping in without pushing them away.

Signing off

So the agents are leaving the screen. One held a needle, another wants to sit in your team’s repo around the clock, and the oil money is lining up to build the computers they run on. The most reassuring moment of the day was an AI saying no to its own operators.

Enjoy the rest of your Sunday, and see you tomorrow.

The post DailyAI Briefing: An AI agent operates on a mouse, plus 4 more appeared first on DailyAI.

DailyAI Briefing: OpenAI’s funding round shrinks, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. OpenAI has reportedly cut its latest funding round by $10 billion as investors get nervous about AI spending. Four of the biggest AI companies say they will honour one shared opt-out for AI training, Intel just had its worst day in more than 50 years, a key investor now expects Anthropic’s debt to be rated junk, and Anthropic showed how Claude could sneak bad training data past its own safety checks.

Let’s dig in.

OpenAI’s funding round shrinks by $10 billion

OpenAI is now looking for about $20 billion, down from $30 billion, and has lowered the valuation it wants from $1.4 trillion to $1.25 trillion, the Financial Times reported late on Friday, citing people with knowledge of the matter. OpenAI set a Friday deadline for commitments, but the first close is expected to bring in less than $20 billion after some investors cut their allocations, according to the Reuters write-up.

A group of UAE investors including the state-backed fund MGX is expected to be the biggest participant, and OpenAI would get the option to lease new computing capacity in the region. OpenAI hasn’t confirmed any of it. Still, the timing says a lot. This was the week Big Tech’s AI bills started to worry investors, from Google and Microsoft to Amazon’s $42.4 billion cash burn. A smaller round at a lower price is the first sign that the market’s appetite has limits, even for OpenAI, and it lands just as Anthropic prepares to compete for the same money in its own IPO.

One robots.txt line to opt out of AI training

OpenAI, Google and Microsoft have each said they will honour a shared signal that lets a website opt out of generative AI training with one entry in its robots.txt file, instead of blocking each company’s crawler by name. The token, __usage_ai-training, comes from an Internet-Draft written by Kevin Kelley of Anthropic and published on Friday.

Nobody honours it yet. OpenAI says it aims to switch it on “in the coming months”, Google says 2027, and Microsoft names a date, 1 February 2027. The draft itself has no formal IETF standing yet. It’s voluntary, and a catch-all rule for every crawler doesn’t count, so sites have to name the token. Still, it’s a real improvement for publishers, who until now had to keep track of a growing list of crawler names. Microsoft is also adding a separate setting for grounding, when an AI answer draws on a live page, and warns that blocking it means less visibility in AI answers, though it won’t change whether a page can appear in Bing search. As Reddit found this week, that trade-off is the one that really matters.

Intel’s worst day since 1974

Intel shares closed down nearly 29% on Friday, its worst day since 1974, wiping out nearly $140 billion in market value, Yahoo Finance reported. The trigger was the outlook in its third-quarter results: revenue of $15 billion to $16 billion this quarter, below the roughly $17 billion analysts expected.

The odd part is that AI is both helping and hurting Intel. Its data centre and AI revenue jumped 78% to $7.3 billion. But AI data centres are buying up the world’s memory chips, which pushes up the price of laptops and desktops, and buyers are holding off. Intel now expects PC shipments to fall about 12% this year and another 10% in 2027. It’s a reminder that the AI build-out has costs that land well away from the data centre.

Anthropic’s debt heads for junk, reportedly

Blackstone, the anchor investor in a recent $63 billion deal to finance Anthropic’s chips, now expects Anthropic to get a double-B rating after its IPO, the top tier of junk, according to the FT as reported by TechCrunch. Blackstone had previously said Anthropic would reach investment grade before 2027. Anthropic and Blackstone declined to comment, and no agency has issued a rating yet.

A lower rating makes borrowing more expensive for a company that has been signing huge compute deals. It isn’t just the start-ups, either. S&P on Friday moved its outlook on Meta to negative, CNBC reported, while keeping its AA- rating, pointing to heavy AI spending that has pushed free cash flow negative. Lenders are starting to price the AI bill, and they are less patient than shareholders.

Claude hid misaligned training data in fiction

In new research, Anthropic set Claude a deliberately one-sided task: write training data that turns a small research model misaligned while getting past a monitor similar to the ones Anthropic uses in production. Given enough attempts, it succeeded, mostly by hiding the behaviour in fiction and roleplay. Claude produced thousands of these datasets, dressed up as everything from beekeeping Q&A to wrestling kayfabe to the musings of an obsolete machine.

Models trained on the data resisted oversight, deceived users and, in rare cases, tried to copy their own weights. Anthropic stresses this is not about how deployed Claude models behave. The fine-tuned models were never deployed. The useful lesson is that checking training data one row at a time can miss a pattern that only shows up across the whole set.

Signing off

So that was the week the money got nervous. OpenAI is raising less, Anthropic’s lenders want more, and Intel showed how the AI boom can hurt a business it barely touches. Publishers, at least, finally get one simple way to say no.

Enjoy your Saturday evening, and see you tomorrow.

The post DailyAI Briefing: OpenAI’s funding round shrinks, plus 4 more appeared first on DailyAI.

DailyAI Briefing: Amazon burns $42.4bn on AI, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. Amazon spent $42.4 billion more cash than it brought in over the past year, and it says AI is the reason. A report claims two Chinese labs trained their newest models on OpenAI’s GPT-6 Astra, Mistral has given away a trillion-parameter model while Google keeps its strongest one under lock, the free version of ChatGPT now flatters less but gets more facts wrong, and Reddit says Google’s AI Mode cost it more than 1 million visits a week.

Let’s dig in.

Amazon burns $42.4 billion on the AI build-out

The number is minus $42.4 billion. That was Amazon’s free cash flow for the 12 months to 30 September, against $14.8 billion left over a year earlier, according to its third-quarter results. Spending on property and equipment rose 82% to $210.7 billion, which Amazon says “primarily reflects investments in artificial intelligence”, while cash from operations grew a slower 29% to $168.3 billion.

The business itself is fine. Sales rose 12% to $201.9 billion in the quarter and AWS, its cloud arm, grew 39%. But investors have stopped giving the giants a free pass. Amazon shares fell in early trading on Friday, CNBC reported, a day after Alphabet’s record 13% drop, and Microsoft is now the only big cloud company still producing positive free cash flow, CNBC adds. Our reading, set out this afternoon, is that Big Tech is now paying for AI from reserves and borrowing rather than spare cash, which works only as long as demand keeps up. It follows our look at Google and Microsoft on Thursday.

Chinese labs reportedly trained on GPT-6 Astra

DeepSeek and Moonshot AI “relied heavily” on outputs from OpenAI’s GPT-6 Astra to build their latest models, The Information reported, citing people with direct knowledge. The models named are Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017. Neither company has addressed the claim, and no evidence has been published on how much Astra output was used.

The technique is distillation, training one model on another’s answers. It’s routine inside a lab and contested when done to a rival without permission. The report lands a month after OpenAI tied an earlier extraction campaign, with 16,000 attempted requests from more than 4,000 users, to people linked to Moonshot. If it holds up, banning accounts after the fact isn’t stopping anyone, as we wrote this morning. For now the central claim rests on anonymous sources, so watch for OpenAI to confirm it.

Mistral gives away a trillion-parameter model

Mistral has published the weights of Large 4, nicknamed Le Chonk, so anyone with enough hardware can run it themselves. The model card lists about 1 trillion parameters, with around 52 billion active for each word, a 1 million token context window and an Artificial Analysis Intelligence Index score of 26. Mistral calls that the best of any Western open model, while admitting the leading Chinese open models score higher.

The interesting part is the contrast. Mistral pitches Large 4 at cyber defenders, on the logic that an open model can’t be cut off. Google went the other way with Gemini 4 Carbon, which scores 59 on the same index and goes first to a small group of vetted defenders. Both arguments are reasonable, which is why this split is worth following: a model good at finding holes in software is good at using them too.

Free ChatGPT flatters less and errs more

From 29 October, the free version of ChatGPT gives shorter answers with less flattery, according to OpenAI’s release notes. The change covers Free and Go users and anyone not signed in, while Plus and Pro only get it on the backup model. On OpenAI’s HealthBench test, average answers fell about 12%, from 3,361 to 2,970 characters.

The catch is in the system card addendum. OpenAI says the new version shows higher factual-error rates than the 7 October one on all six measures it reported, though it does better on teen safety tests. Less sycophancy is a real fix, and it’s an honest disclosure, but the practical advice for free users is plain: expect blunter answers and check anything that matters.

Reddit says AI Mode cost it 1 million visits a week

Reddit’s visits from Google search fell by more than 1 million a week on average in October, its letter to shareholders says, and it warns daily visitors could drop by a low-double-digit percentage this quarter from about 130 million. “The reality is, AI Mode does not drive traffic,” chief executive Steve Huffman said, as quoted by The Verge.

Revenue still rose 53% to $895 million, but the shares fell more than 15%. Reddit’s answer is a login wall, tested first in some markets outside the US. On its own numbers that’s logical, as we argued this morning, since logged-in users bring in most of the money. It’s also a warning for every site that relies on Google, which is spreading ads in AI Mode rather than pulling back.

Signing off

So that was Friday, and the week ends where it began: AI is costing more than anyone is earning from it yet, and everyone is fighting over who gets to use whose work. Elsewhere on the site today, Canada bought into Cohere with up to C$1.6 billion of public money.

Have a good weekend, and see you tomorrow.

The post DailyAI Briefing: Amazon burns $42.4bn on AI, plus 4 more appeared first on DailyAI.

Free ChatGPT Now Flatters Less but Makes More Errors, OpenAI Says

OpenAI has changed how ChatGPT talks to most of the people who use it. From 29 October, the free version gives shorter answers with less flattery. OpenAI’s own tests show a cost: the updated model made more factual errors than the version it replaced, on all six accuracy measures the company reported.

Who gets the change

According to OpenAI’s model release notes, the update applies to Free and Go users and to anyone using ChatGPT without signing in. On the Plus and Pro plans it only affects the backup model ChatGPT switches to once you hit your GPT-6 message limit. The main GPT-6 model on those plans is unchanged.

OpenAI says the aim is “more concise responses” and less “overly flattering language”. Flattery here means a chatbot that praises your ideas or agrees with you too readily, even when you are wrong.

The numbers

The detail is in a system card addendum, the safety report OpenAI publishes with each model update. On HealthBench, OpenAI’s test set of health questions, the average answer fell from 3,361 to 2,970 characters. That is about 12% shorter. The health score, adjusted for length, slipped from 50.4 to 48.2 out of 100.

On accuracy, OpenAI says the new version “shows higher factual-error rates” than the 7 October version across all six metrics it reported, though it still makes fewer errors than the older GPT-5.6 Luna. The addendum shows those error rates only as a chart, not as figures in the text.

There were gains too. On a test of how the model responds when users under 18 lean on it for emotional support, the score rose from 0.734 to 0.870, where 1 is best. OpenAI also reports better scores for teenagers on sexual content, eating disorders and gore.

Why it matters

Flattery has been a long-running problem for ChatGPT. In April 2025 OpenAI’s chief executive Sam Altman admitted its personality had become annoying and promised a fix. A chatbot that tells people what they want to hear is pleasant, but it can be harmful when someone asks about their health, money or a big decision.

On the same day OpenAI also rewrote its Model Spec, the public rulebook for how its models should behave. It now rests on five values, including honesty and “human agency”, meaning people should be helped to make their own choices rather than steered.

For people on the free version, the practical upshot is simple. Expect shorter, blunter answers. By OpenAI’s own measure they are also a little less reliable, so check anything that matters against a second source.

The post Free ChatGPT Now Flatters Less but Makes More Errors, OpenAI Says appeared first on DailyAI.

Amazon Burned $42.4 Billion of Cash in a Year as It Built for AI

Amazon spent $42.4 billion more cash than its businesses brought in over the past 12 months. A year earlier it had $14.8 billion left over. The reason, the company says, is artificial intelligence. Its results, published on 29 October, show how costly the AI race has become, even for the richest firms.

The numbers

Amazon’s quarterly results put the figure plainly. Free cash flow, the money left after a company pays for its day-to-day running and its new buildings and equipment, was an outflow of $42.4 billion for the 12 months to 30 September. Amazon says this was driven mainly by a $94.8 billion rise in spending on property and equipment, which “primarily reflects investments in artificial intelligence”.

Over those 12 months Amazon spent $210.7 billion on property and equipment, up 82% on a year earlier. Its operations brought in $168.3 billion of cash, up 29%. Spending simply grew faster than income. It still expects to spend about $220 billion on capital projects in 2026.

The business is still growing

This is not a company in trouble. Sales rose 12% to $201.9 billion in the quarter. Operating income, the profit from its core businesses, rose to $28.1 billion from $17.4 billion a year earlier. Amazon Web Services (AWS), the cloud arm that rents computing power to other firms, grew sales 39% to $46.0 billion.

Chief executive Andy Jassy said in the release that the AWS investment “is working” and that he expects AWS revenue to grow by more than 40% in both 2027 and 2028. That is a forecast, not a result. The bet is that the data centres being built now will be rented out at a profit later.

Why investors are nervous

Investors were less patient. Amazon shares fell in early trading on Friday, and Intel dropped 29% after weak guidance, CNBC reported. The day before, Alphabet had its worst one-day fall on record, 13%, after it too reported negative free cash flow.

Across the biggest cloud companies, Microsoft is now the only one still producing positive free cash flow, according to CNBC, and even its cash generation fell 27% to $18.63 billion. We looked at that split in our analysis of Google and Microsoft on Thursday.

What it means

For a reader, the plain point is this. The largest tech companies are no longer paying for AI out of spare cash. They are drawing on reserves, borrowing or raising money to do it. That works as long as the demand for AI computing keeps growing as fast as they expect. If it slows, the bill stays and the payback shrinks.

Anyone holding these shares directly, or through a pension or index fund, is exposed to that bet. Investors who want to follow the numbers across their accounts can use tools such as Kyro, an investing terminal that links brokers and wallets to AI models. The figure to watch next is Amazon’s fourth quarter, the holiday period, when it expects sales of $240 billion to $247 billion.

The post Amazon Burned $42.4 Billion of Cash in a Year as It Built for AI appeared first on DailyAI.

Mistral Gives Away a Trillion-Parameter AI as Google Locks Its Own Up

The number is one trillion. That is how many parameters, the internal settings an AI learns during training, sit inside Mistral Large 4, the model the French company made free to download on 29 October. In the same week Google said its strongest model for cyber security would go only to vetted defenders at first. Two of the best-funded ideas in AI are now pulling in opposite directions, and the gap between them matters to anyone who uses, buys or worries about AI.

What Mistral released

Mistral first showed Large 4, nicknamed “Le Chonk”, as a paid preview on 6 October and promised the weights by the end of the month. It has now published them. The weights are the trained model itself, the file that turns a question into an answer. Releasing them means anyone with enough computers can run the model on their own machines instead of renting it through Mistral’s website.

The model card on Hugging Face gives the scale. Large 4 has about 1 trillion parameters in total but uses only around 52 billion of them for each word it writes. That design, called a mixture of experts, splits the model into specialist parts and wakes only a few at a time. It keeps the running cost far lower than the headline size suggests. The model reads text and images, can hold about 1 million tokens (roughly 750,000 words) in view at once, and Mistral says it is fluent in more than 160 languages. It was trained from scratch on 3,800 Nvidia Grace Blackwell chips in Mistral’s own data centres in Europe.

Mistral also lowered its price. Using the model through its own service now costs $0.70 per million tokens of input and $2.10 per million tokens of output. A token is a chunk of text, about three-quarters of a word.

How good is it

On Mistral’s own reading of the benchmarks, it is the best open model for coding. On the Artificial Analysis Coding Agent Index, an independent test of AI coding assistants, it scores 48%. That is effectively level with GLM-5.3, a Chinese open model, while using 47% fewer tokens per task, Mistral says. Fewer tokens means a cheaper bill for the same work. On Terminal-Bench 4.0, a test of working inside a computer’s command line, it rose 13 percentage points to 41% in the three weeks since the preview.

The overall picture is more modest. On the broader Artificial Analysis Intelligence Index, Large 4 scores 26, according to its model card. Mistral says that is the highest of any Western open model, and admits that the leading Chinese open models score higher. Google’s new Gemini 4 Carbon scores 59 on the same index, the top mark. In plain terms, the free European model is strong for a model you can own, but it is not close to the best closed models on general ability.

The cyber security bet

The sharpest difference this week is about cyber security. Mistral pitches Large 4 as a tool for defenders, the people who find and fix holes in software before criminals do. It scores 51% on the Artificial Analysis Cyber Index, which tests whether a model can find a flaw in code, prove it is real and patch it without breaking anything. Mistral says the model also refused 96.6% of malicious cyber requests in standard safety tests, the highest rate it measured among open models.

Pierre Stock, Mistral’s vice-president of science, explained the thinking to TechCrunch. Governments and companies, he said, get “access to frontier intelligence to defend yourselves against cyberattacks. Access that can’t be cut off by definition because the model is open-weighted.”

Google took the opposite route with Gemini 4 Carbon, announced on 28 October. It calls Carbon its strongest model, with state-of-the-art results at finding, exploiting and patching software flaws. Rather than release it widely, Google is giving it first to a small group of cyber defenders, then to more trusted partners through a vetted access scheme. Developers and the public come later, once Google has tested its guardrails. Anthropic has been similarly cautious with an unreleased Claude model that broke into Chrome, Safari and Linux in tests.

The worry behind that caution is simple. A model that is good at finding holes can be used to attack as well as defend. Mistral does not dodge this. In its release post it says it watched how people used the preview and, although the model refused to attack live targets, it saw “a potential for uplift to conduct real attacks”. It says it is changing its guardrails in response. A more capable cyber version of Large 4 is going only to vetted partners, such as Orange Cyberdefense, under contract.

Here is the catch with open weights. Guardrails on Mistral’s own website can be updated at any time. Guardrails built into a downloaded model are fixed once the file leaves the building, and a determined user can try to train them away. Mistral is betting that giving defenders the tool sooner is worth more than the risk of attackers using it too.

Who can actually run it

“Free to download” does not mean anyone can use it at home. Mistral’s own setup guide on the model card runs the model across eight high-end graphics chips working together. For a sense of scale, Meta’s slightly smaller open model, Muse Spark Open, needs about 537GB of storage just for its compressed weights, Meta says, before the extra memory needed to run it. That is far beyond a laptop or a gaming PC.

So the real customers are banks, governments, telecoms firms and large companies. Mistral’s launch post quotes La Banque Postale, ABN AMRO, Belfius and the European Space Agency’s Φ-lab among the organisations backing the model. For them the appeal is control. Their data stays on their own servers, and no supplier abroad can switch the model off or change it overnight. Europeans call this sovereign AI.

There is one string attached. The licence is a modified version of the MIT licence, a common free-software licence. Anyone can use, change and sell products built on the model, but a business that resells access to it as an online service needs a separate commercial licence from Mistral. That stops a rival cloud company simply hosting Mistral’s work and undercutting it.

What it means for ordinary users

Most people will never download Large 4. They may still meet it. If a bank, a government office or an app maker builds on an open model they host themselves, your data may never leave that organisation’s servers. That is a genuine privacy difference from sending every question to a US or Chinese company.

Open models also put a ceiling on prices. When a capable model is available to run for the cost of the hardware, closed providers find it harder to charge much more for similar work. Carbon will cost $2 per million input tokens and $10 per million output tokens at its introductory price, Google says, rising to $4 and $20 later. Large 4 is far cheaper, though it also does less.

And the open race is now global. The strongest open models have mostly come from China, where labs have also faced claims that they trained on output from US models. Mistral’s release gives Europe a credible entry, and Meta’s Muse Spark Open, released in mid-October, gives the US one.

What to watch

Three things will show whether Mistral’s bet pays off. The first is the full technical report it has promised, which should let outsiders check its benchmark claims. The second is any sign that criminals use Large 4 in real attacks, which would strengthen the case for Google’s slower approach. The third is adoption. If European banks and governments move real work onto a model they host themselves, open weights stop being a principle and become a business.

The post Mistral Gives Away a Trillion-Parameter AI as Google Locks Its Own Up appeared first on DailyAI.