Every story, at length

Jakob Macke, Daniel Gedon and Stefan Wahl, writing in a guest post for Anthropic, set out how AI is changing three parts of scientific work: building a model, fitting it to measurements and interpreting what it says about the real world. Their central warning is that a model which reproduces the data is not necessarily an explanation. Many different models can fit the same results, and an AI agent that returns just one hides the alternatives.

The more established advance is in fitting models. A method called simulation-based inference runs a computer model repeatedly with different settings, then trains a neural network to work backwards from an observed result to the range of settings that could have produced it. The simulations can take weeks, but once trained the network can analyse new data in seconds. The approach has been used in cosmology, structural biology and geophysics; when two neutron stars collide, it can estimate their location quickly enough for telescopes to be pointed while the signal is still detectable.

The newer step is letting AI design the model itself. Systems appearing over the past two years use a large language model to write candidate models as code, test them against data and revise them. The researchers argue that these systems should also produce a spread of plausible models, with weaker candidates discarded and new variations built on stronger ones. Features shared by all the survivors are likely to matter; disagreements show where another experiment is needed. But the evidence is thin where it counts most: these systems have mostly been tested on rediscovering models scientists already know, with sparse evidence on unanswered problems.

The historical comparison makes the stakes clear. In 1951, Andrew Huxley spent more than three weeks solving nerve-cell equations on a hand-cranked calculator after Cambridge University's computer went out of action for six months. An AI system could plausibly find a similar equation in an afternoon. Yet Hodgkin and Huxley's crucial insight came from knowing that simpler versions failed: their fourth-power term suggested four moving parts in a cell-membrane gate, a hypothesis later supported by protein structure. Interpretation remains the hardest step to automate, because there is no clear way to measure whether an AI system understands a system rather than merely fits its data.

Anthropic has published research on how its Claude agents retain knowledge between sessions, using notes, searchable records and a process it calls dreaming. During dreaming, the agent uses idle computing time to review earlier work, check notes against original transcripts and reorganise what it has learned. The findings matter to businesses deploying agents for long tasks: persistent memory can help, but indiscriminate note-taking consumes time and can preserve errors.

The clearest example came from Pokémon Red, which Anthropic used to test judgment over extended tasks. About 10,000 steps into separate runs, Claude Sonnet 3.7 had written 413 memory files across 225 folders and earned one badge. Claude Fable 5 had written nine files in two folders, plus one overview file, and earned four badges. The older model recorded battles, items and other details that were often visible on screen or quickly became outdated. At one point it spent 17 steps creating and deleting memory files without taking any action in the game. Fable 5 instead kept plans, strategies and tested rules. Anthropic says Opus 5.5 later used a similarly small set of notes and beat the game.

The research also challenges a natural instinct for companies building with agents: prescribing exactly what they should remember. On a database-question benchmark, telling Opus 5 and Opus 4.7 to record only failures removed much of memory’s performance benefit. Newer models generally performed better when allowed to judge what was worth saving. Older models needed more direction; instructing Sonnet 4.6 to read its notes and test its theories nearly doubled its benchmark score.

Dreaming supplied a further improvement by auditing memory after a session. In one Pokémon run, Claude recorded a landmark on the wrong side of a mountain and repeatedly took a wrong turn. A dreaming pass compared the note with the transcript, corrected it, and the agent subsequently reached its destination. Anthropic also tested agent-built wikis on legal and company documents. They helped when answers required information from several documents, but offered no clear gain when a simple text search sufficed; on one task, the wiki grew larger than the underlying documents. The genuinely new contribution is this set of evaluations and design lessons, rather than a new memory product. Anthropic says open questions remain about computing costs, update frequency and preventing accumulated errors.

ToyBlocks is a simulated tabletop with a robot arm and lettered blocks. Researchers can automatically generate tasks and training demonstrations, then test skills such as pushing, flipping and stacking. It is the first of Allen AI’s planned FoundationWorlds: controlled settings intended to make robot experiments repeatable and failures easier to trace. The release includes simulation tools, training and evaluation code, and two trained models.

The early results are a warning against reading polished demonstrations as evidence of broad competence. Initial tests of ManiFlow, π0.5 and MolmoAct 2 found that these action models generalised poorly despite large-scale pretraining. The experiments varied block positions, stack heights, letters and colours, including tasks outside the training range. Allen AI says action-model generalisation remains far from solved. That matters to companies developing warehouse, manufacturing or domestic robots: performance in a familiar setup may not predict what happens when the layout or task changes.

The most striking finding concerns a Claude-based agent asked to build towers spelling three-letter words. Half the scenes lacked a required letter, so the correct response was sometimes to declare the task impossible. It solved the task or correctly reported infeasibility in 63% of cases; errors included confusing look-alike letters and choosing the wrong orientation. When the agent ran small experiments and rewrote the descriptions of its own skills, success rose from 24% with default descriptions to 63%, beating 36% with descriptions written by a human expert.

The framework also produced evidence that simulation can transfer to physical machines. ManiFlow’s individual skills succeeded 80.0% of the time in simulation and 81.1% on a real robot arm. A combined Molmo 2 and ManiFlow system completed 50.0% of real-world block-spelling episodes, against 55.5% in simulation, though Allen AI says more trials are needed. Separately, adding 100 trajectories generated by the best-performing video model raised a policy’s success in an unseen region from 19% to 78%.

The genuinely new contribution is the controlled, open testing and data-generation system, rather than a claim of a generally capable robot. The initial release contains 85,336 long-horizon demonstrations, covering nearly 1,600 hours of robot execution. A larger dataset, more skills and broader benchmarking are targeted for the coming weeks.

Spaces are breakout conversations organised around a particular topic or goal, living in the sidebar of the interface where users interact with their Muse agent. Clicking on a Space opens the conversation, where Muse works toward that goal on the user's behalf. Meta describes them as "custom experiences" and says users can describe an idea in chat and Muse will build a space for it, whether that is a tracker, a planner or a project board. Examples the company gives include planning a vacation, collecting recipes for a family event, tracking holiday gifts, planning a wedding or managing DIY projects.

The more consequential addition is integration with third-party apps and information. Meta says that once users connect the apps they use, their spaces keep themselves up to date in the background, so new concert dates, price changes or savings progress appear without the user having to ask. That moves Muse beyond a chat window that waits for instructions and toward something that monitors connected services and maintains a live view of a project. For businesses building consumer apps or services, it is a signal that Meta wants its agent interface to sit between users and the apps they already rely on.

The feature is the second Muse addition Meta has announced in the past week. On Thursday, the company introduced Muse Voice, which lets users video chat with their agent and give it commands verbally. Mark Zuckerberg, Meta's founder and chief executive, posted on X on Friday that he had been playing with spaces in Muse and had made a barbecue game, an event tracker and a home control dashboard.

Meta is building on early traction. During its earnings call in October, Zuckerberg said the agents had reached 10 million users in the United States and Canada, less than two months after they debuted. The company is competing with other cartoon-styled AI agents, including OpenAI's dots and SpaceXAI's Grokbot, which launched with similarly playful personas. Spaces itself is an incremental feature — persistent project threads are not a new idea — but combined with app connections and voice, it shows Meta moving quickly to turn an early user base into a more capable assistant.

Amazon said on Friday that customers invited to the service will receive a text asking them to enable it. They can then message Alexa in ordinary language to find products, add them to a cart and complete a purchase on Amazon. Others can join a waitlist, with a broader rollout planned in the coming weeks.

The service extends Alexa's recent shopping upgrades rather than starting from scratch. It can manage lists, track prices, search for items across the web, compare products and check order status. Shoppers can also send a picture of an item, or of someone wearing an outfit, to find similar listings on Amazon. The company says Alexa will learn users' preferences over time and suggest products accordingly.

The timing exposes the commercial tension behind consumer AI agents. Amazon blocked Meta's Muse from querying its site in September; it is now offering its own text-based route into shopping, similar to agents such as Muse and Instinct. For retailers and brands, that raises the possibility that an AI assistant controlled by the platform becomes another gatekeeper between shoppers and product listings. For Amazon, it creates a direct way to keep product searches, price comparisons and purchases inside its own service.

The most striking evidence that this market is still early comes from DoorDash. The delivery company said last month that third-party agents had placed only hundreds of orders in total on its platform so far. It plans to concentrate on its own text-based agent, which launched in September. That is a small base compared with the attention consumer agents are attracting, and it suggests the immediate contest is over who owns the customer relationship, not over a proven surge in agent-driven sales.

The game presents a retro-themed room containing a typewriter, a CRT monitor, a desk lamp and a music tuner. Users can type text on the typewriter and have it spoken back, or rotate the tuner to generate different kinds of music. The lamp also makes a sound when its switch is clicked. An introductory warning about “intrusive” cloud providers in the town is an apparent dig at cloud-based AI models.

The demonstration is powered by Woof, which Underdog says can run on most devices without an internet connection. The model is 32MB in size, has 31 million parameters and is open for anyone to download. Parameters are the internal settings that determine how an AI model behaves; the important point for non-specialists is that this system is designed for local use rather than relying on remote servers.

That could make the technology useful for local assistant experiences or personalised podcasts, although the source cautions that such a small model is unlikely to replace larger systems. For businesses and developers, the attraction is the possibility of adding speech features without depending on a cloud provider. For users, running the model on their own device offers a route to keeping their data under their own control.

Woof follows Bark, a 167MB speech-to-text model that Underdog launched weeks earlier for transcription and dictation on edge devices such as phones and computers. Founder Sigil Wen told TechCrunch that he began building the app because he wanted an alternative to notetakers such as Granola, using AI models that run on users’ devices and let them own their data. He said he expected “most capable models of today” to run on devices people already own within months, and described Underdog’s goal as training capable, open, American-made models that enable private AI.

The company has also released a 27-billion-parameter model that can run on laptops. Separately, it has a waitlisted app positioned as a locally hosted alternative to Muse and Instinct. The genuinely new element here is the extremely small offline speech generator and its interactive browser demonstration; the broader effort to move AI processing from the cloud onto personal devices is already being pursued by other researchers and startups.

The new material comes from Liva AI and covers spontaneous, conversational speech rather than scripted readings. The Egyptian set runs to 1.34 hours and the Nigerian set to 1.55 hours, with transcripts that keep punctuation, capital letters and disfluencies — the hesitations and false starts of ordinary talk. Both are private evaluation sets: the audio is not published for model makers to train on.

That privacy is the point. A few months ago the leaderboard added 11 private splits from Appen and DataOceanAI covering scripted and conversational speech in five accents, and it has since added an Indian-accented English split with Voice Arena. The organisers say the new data is meant to minimise "benchmaxxing" — tuning a system to do well on a known test rather than on speech it has never seen. For a business buying transcription or voice-agent software, a ranking built on held-out recordings is a more useful signal than one a vendor could rehearse for.

The leaderboard is also changing how it reports results, grouping accents using Kachru's Three Circles, a framework credited to Hume AI's leaderboard, and promising to split the Outer and Expanding Circles as more accents arrive. The aim is a breakdown that shows where a model works and where it does not, instead of a single English score that can hide weak performance on speakers outside the US and Britain.

The genuinely new part is narrow: two accents, under three hours of audio between them, and one more data provider. The incremental part is the continuation of a stated push for more accents, more languages and more private providers. Liva AI says it collects multimodal interaction data through consumer products used by real people, and that owning both collection and distribution lets it build products around specific model gaps, yielding natural behaviour rather than task-based interactions. No model rankings or accuracy figures were published with the announcement, so there is no evidence yet of which systems gain or lose from the tougher tests.

Crusoe confirmed on Friday that it had sold the project to Google. The search company had reportedly raised concerns about the development’s cost and construction timeline under Crusoe’s management. Google did not disclose the purchase price, but said it planned to invest $17 billion in developing the data center campus.

The project was announced last year by Crusoe and Tallgrass, an infrastructure company backed by Blackstone, with plans eventually to reach 10 gigawatts of power capacity. Its change of ownership shows the difficulty independent developers face in delivering very large data centers on time and within budget. Securing equipment and power is challenging, and opposition from local communities can add further complications. For businesses buying AI computing capacity, the deal is a reminder that announced campuses do not necessarily become available on the timetable first promised.

The most striking detail is that Google has moved from prospective customer to owner after Crusoe failed to land its business. That is a setback for Crusoe’s role on this project, but not a complete break between the companies: Crusoe is still developing another data center for Google near Amarillo, Texas. Bloomberg first reported the Wyoming sale in June, and Crusoe went on to raise $3.9 billion in September despite the setback.

Crusoe, an eight-year-old Denver-based company recently valued at $30.9 billion, leases data center space to customers that bring their own AI chips, rents out its own chips and sells computing power for running AI models. Its largest projects serve major technology groups. Its Abilene, Texas campus is being built in phases for Oracle and OpenAI, while a separate 900-megawatt campus next door is being constructed for Microsoft, with the first building expected to come online in mid-2027. Meta is also a customer, and Bloomberg reported that trading firm Jane Street signed a $13 billion, five-year cloud deal with Crusoe. The genuinely new development is Google’s takeover of the Wyoming site; Crusoe’s broader business and other projects remain in place.

A new report from Ramp shows AI spending flatlined over late summer even as usage volumes rose. Token spend grew just 1.5% across August and September, compared with 42% growth in July from the previous month. October brought a partial rebound, with token spend up 8.9% from September 23 to October 23, but Ramp's lead economist Ara Kharazian said the headwinds on spending growth remain strong.

The slowdown is not a demand problem. Kharazian points instead to price competition: OpenAI cut API prices in half for GPT-6 Sol and GPT-6 Luna in September, and Anthropic made significant reductions to pricing for Fable 5.1 and Opus 5.5, a decision the company said in a recent filing slashed revenue growth by about $600 million. Both companies have also offered enterprise discounts intended to lock in spending that are instead hitting their revenue targets. Customers are getting thriftier too, setting budgets for employees and routing work to cheaper models. Notion co-founder Akshay Kothari posted internal data matching Ramp's findings, summarised as: "tokens go brrr. costs go ahhh."

The concentration of spending makes the slowdown sharper. Almost all AI spending comes from venture-backed, high-growth tech companies, and the top 1% of firms have stuttered slightly, falling just below the $10,000 per employee they were hitting at summer peaks. Median firms spend $12 to $15 per employee per month, and the top 10% spend $600 to $1,000. Kharazian said a slowdown among that top cohort is sufficient to slow revenue entirely despite rising volumes, and it is not clear other industries will ever reach that level: a doctor's office or car dealership is unlikely to need to spend that much.

The timing is awkward for Anthropic, which is gearing up for an IPO. Bloomberg reported Friday that its potential valuation sank from the $2 trillion estimated in early October to $1.2 to $1.3 trillion, with a high junk rating expected from credit agencies. OpenAI lowered its proposed valuation to $1.25 trillion from $1.4 trillion and cut its latest fundraising goals by about $10 billion. Investors hope new products such as personal agents will drive the next wave of spending, but Kharazian is skeptical they are as economically productive as coding agents, which emerged just a year and a half ago as the most commercially viable AI product.

Censys, a cybersecurity firm that scans the internet for exposed systems, identified the deployments of Artex, an open source tool that automatically probes computer systems for weaknesses. Artex has been linked to several cyberattacks since its public release in July. Microsoft has described it as a system of AI agents that can conduct attacks from reconnaissance through data theft with little human interaction. Because it is free to download and run, it has become popular with people who cannot access Anthropic or OpenAI's most advanced cybersecurity models.

The exposure cuts both ways. An open Artex system could let an outsider use it to attack unsuspecting targets, but it also reveals the operator's own attack tools, including implants — hidden malicious software used to keep access to a victim's network. Silas Cutler, the Censys principal security researcher who wrote the findings, said not every instance is run maliciously and many are used by authorized testers. It is unclear why so many are reachable from the web, since the software does not need to be.

The linked attacks show the stakes for ordinary businesses. Security companies have tied Artex to a September campaign against the Korean finance industry, which CrowdStrike attributed to a Chinese-speaking hacker who had left an open directory hosting the system. Censys said that instance had been publicly exposed for months and appeared to hold attack data reaching well beyond South Korean banks, including payment and gaming companies. Artex was also used in an October attack on IT support giant ConnectWise that let Chinese-linked hackers spy on thousands of organizations whose data and files are hosted in Microsoft's cloud. Microsoft said the hackers broke into customers' cloud environments and stole data so quickly that defenders had little time to detect and respond.

The genuinely new finding is the scale of the exposure, not the existence of AI-assisted hacking. Many of the attacks linked to Artex relied on basic weaknesses such as weak or known passwords and a lack of multi-factor authentication, the common login safeguard that asks for a second proof of identity. Artex has since said it will go closed-source, limiting further releases to paid customers, but dozens of copies of its code remain on code-sharing sites and can still be used in active attacks.

The stolen information includes names, usernames, password hashes and live tokens. Password hashes are scrambled versions of passwords, while live tokens can act like temporary digital keys for an account. Chainalysis said the attackers gained unauthorised access to its systems and internal tools using AI agent-assisted offensive techniques.

The customer list makes the breach especially sensitive. Chainalysis and public sources have identified users including the FBI, US Immigration and Customs Enforcement, the US Air Force, Microsoft, Coinbase and Barclays. The company analyses blockchain activity, so its tools sit close to investigations and financial compliance work. It remains unclear whether the theft exposed customers to further threats.

The company’s account of what was taken has also changed. In an email last Friday, Chainalysis told affected customers that an October cybersecurity incident had exposed what it called “non-sensitive customer data”, including user IDs, full names, email addresses and organisation role labels. The source questions why that information was treated as non-sensitive. In a Wednesday update, Chainalysis said the incident was part of an “ongoing multifront cyberattack” and that a separate on-path attack had exposed a subset of customer passwords.

An on-path attack, also called an adversary-in-the-middle attack, involves intruding into infrastructure that carries traffic between two parties. Chainalysis said that attack route had been identified and fixed, but it has not said where the infrastructure was compromised or whether the captured data was protected by TLS encryption, a standard method for securing data in transit. The genuinely new development is the Wednesday disclosure of password exposure and the claim that credentials and tokens cover all users, beyond the more limited customer-data access described earlier.

The startup’s robots can vacuum and mop without human help, according to cofounder Cornelia Weinzierl, but tasks such as making beds and folding laundry still require a person to guide their movements remotely. Pilots watch through cameras in the robots’ heads and wrists. For some jobs, the operator performs the motions and the robot copies them, generating examples that Tau hopes will teach future machines to work independently.

That makes the service less a finished product than a way to collect training data from real homes. Tau has completed only a couple hundred cleanings so far. Its pricing reflects the experiment: an early offer was $30 for an hour with one robot, while the current charge is $60 for a weekly cleaning with two robots or $100 for a one-time visit. The reporter who tested it found the machines thorough but slow, and concluded that cleaning yourself would be a better use of time than watching them work.

The visit also showed how far the machines are from the effortless domestic robots of science fiction. The four-and-a-half-foot-tall machine staggered up the front steps, was defeated by a carpeted staircase and had to be carried. It later stopped cleaning when hot weather disabled the equipment used by its human controller. On a smoother second visit, it cleaned the kitchen and toilet, but left taps dripping, knocked over a closed olive-oil bottle and could not right an air purifier it had toppled. Its most impressive trick was repeatedly reattaching the head of a toilet-cleaning wand after it fell off.

The genuinely new part is that a startup is putting humanoid cleaners into ordinary homes and charging customers, rather than demonstrating them in a laboratory. The autonomy remains incremental. If Tau or its competitors eventually make cleaning cheap and independent of remote pilots, professional cleaners could lose work. Weinzierl argues that robots could fill jobs marked by high turnover and injury risk, and says Tau is considering other demanding uses such as mold remediation and home health care.