DEV Community: Fathom News

Free Drex Max AI Runs On Your Mac And Outscores Claude Haiku At Decisions

Fathom News — Mon, 02 Nov 2026 04:00:38 +0000

Nace AI has given away Drex Max, a free AI model that picks answers instead of writing them, and it runs fully on Apple M4 and M5 Macs with 24GB of memory or more, so a business can sort tickets and files without sending them to the cloud.

Key Takeaways

Nace AI released Drex Max, a 27 billion parameter model, as free open weights
It reads a document and a question, then gives a probability for every possible answer without writing any text
A 14.45GB version runs on Apple silicon Macs with 24GB of memory or more
Nace says it scores 64 on the Artificial Analysis Decision Index, ahead of Claude Haiku 5.5 (Low) at 61

The company announced the release on X on Oct 30, saying Drex Max "runs 100% locally on Apple Silicon (M4 & M5)." The model and its full spec sheet are on Hugging Face.

What Drex Max Actually Does

Most chatbots write an answer one word at a time. Drex Max does not write at all.

You hand it some context, such as a customer email, plus a question with a fixed list of options. It sends back a score for every option in a single step.

Nace gives a simple example. A customer writes in saying they were charged twice for one order. Asked whether billing, technical or "other" should handle it, the model gave billing a 99.88% chance.

It can weigh up to 255 options per question and read plain text, structured data and images. That makes it a fit for the dull but costly jobs inside AI systems, like routing tickets, flagging risky content or choosing which tool to use next.

Running It On A Mac

The full model takes about 49GB and is built for Nvidia graphics chips. For Macs, Nace made a smaller 4-bit version, a 14.45GB download that runs on Apple silicon machines with 24GB of memory or more.

There is a catch. The Mac version reads text only, so the image features stay on the larger setup. A bigger 8-bit file also runs on Macs but needs 48GB of memory.

Nace says it tested the small version on an M5 Pro Mac with 48GB on Oct 30. Its scores moved by no more than 0.0329 compared with the bigger file, and none of the final picks changed.

How It Stacks Up

Nace lists these scores from the Artificial Analysis Decision Index, which tests how well models pick correct answers. Drex Max ranks fourth of 18, one point behind the leaders.

Model	Type	Score
dl-ultra	Decision	65
SubQ Snap 1.1	Decision	65
Elephant 9B	Decision	65
Drex Max	Decision	64
Claude Haiku 5.5 (Low)	Language	61
Drex 2.0	Decision	58
GPT-6 Luna (Low)	Language	55

On its own Decision Index 0.4, Nace says Drex Max scored 67.52, up from 61.25 for its older Drex v2.0. These figures come from the company, and no outside group has published its own tests of the Mac version yet.

Why It Could Cut AI Bills

AI agents make lots of small choices: which document to read, which tool to call, whether to stop. When each choice goes through a paid cloud service, the costs add up fast.

A free model that makes those calls on a Mac could take a share of that work off the bill and keep private files on the machine. Nace has not published any cost savings figures, so the real gain will depend on the job.

The model comes under Nace's own Open RAIL-M license, which lets people use it for free but sets rules on how it can be used.

What To Watch Next

The test now is whether outside developers get the same results on everyday Macs with messy real-world files. If they do, small local models like this could become a standard first step in AI tools, with the cloud saved for the heavy work.

Originally published on Fathom News.

Anthropic Website Slip Reveals Secret Claude Mythos 5.5 Built For Hacking And Biology

Fathom News — Mon, 02 Nov 2026 00:45:32 +0000

AI company Anthropic briefly posted text on its site pointing to Claude Mythos 5.5, a stronger chatbot for hacking and biology work, hinting at a riskier product for big customers.

Key Takeaways

Anthropic appeared to reference Claude Mythos 5.5 on its website Sunday
The X account TestingCatalog posted the find and included the quoted line about Claude Mythos 5.5
On Friday Anthropic dropped the five-hour cap on Claude Ultra for verified users
Anthropic has not publicly announced Claude Mythos 5.5 in its news feed

The X account TestingCatalog posted the find, including the quoted line "Claude Mythos 5.5 - Our most capable model class for cybersecurity and biology." The post did not show Anthropic shipping the model, only that the text was present on an official Anthropic page.

Anthropic builds the Claude family of AI models and has increasingly separated products by safety profile, access tier and enterprise use case. That matters because a named "most capable" class for cyber and biology would imply a more explicit product split than the company has shown in public releases.
Its SEC filing outlined how the company plans to sell different model capabilities while managing higher-risk deployment.

Claude Mythos 5.5 Points Past Chatbots

The wording matters more than the version number. "Model class" suggests Anthropic may be grouping several systems or configurations under one label rather than naming a single checkpoint. That can mean the same base model ships with different tool access, rate limits, retrieval settings or policy filters.

Cybersecurity and biology are the two domains where frontier model labs face the most scrutiny because of dual-use risks, a model that helps write code can also help find software flaws, and a model that summarizes biology papers can speed harmful research workflows.
In this case, the website reference explicitly pairs the Mythos label with those two domains, signaling Anthropic is considering domain-specific packaging for higher-risk tasks.

Also Read: Anthropic Study Shows Hidden AI Advice Sways Experts 86% Of Time

From Claude Ultra To A More Explicit Ladder

The new text fits a broader pattern in Anthropic's pricing and access changes during the past two days. On Friday, the company dropped the five-hour cap on Claude Ultra for verified users, widening access at the top of the stack.
That move focused on consumption rules. The apparent website string focuses on capability labeling.

Access tiers tell customers how much they can use a model.
Capability tiers tell them which work the vendor believes each model should handle.

If Anthropic is building a separate Claude Mythos 5.5 layer, it may be trying to give regulated buyers a cleaner procurement story. Government buyers, defense contractors and large security teams often want clearer boundaries about which model can touch sensitive code, which logging controls apply and whether a version sits behind extra screening.

What Anthropic Has Not Confirmed

Anthropic has not publicly announced Claude Mythos 5.5 in its news feed, and the X post alone does not prove a release date, price or customer list.
Website artifacts can appear before launch or survive internal testing without becoming public products. Still, hidden references matter because model labs often wire pricing pages, developer docs and policy text into public sites before a formal reveal.

The next thing to watch is whether Anthropic adds documentation, API naming or pricing pages that repeat the Mythos label.
If it does, the story shifts from a leak to product confirmation. If not, Sunday's find still shows where Anthropic's safety and enterprise packaging appear to be heading.

How Anthropic Could Package One Risky Model

A "model class" like Claude Mythos 5.5 would let Anthropic sell the same core reasoning system in several wrappers rather than train a separate model for every buyer.
For example, Anthropic could pair one Claude family model with extra code tools for a security team while giving a biology customer stricter file limits, slower outputs or more aggressive refusals.

That packaging matters for customers like Carahsoft buyers and federal teams because procurement often depends on operational controls, not just benchmark scores. Read Next: Crypto Hackers Stole $6.48B In October, But One Heist Was 97% Of It

Originally published on Fathom News.

Saudi Oil Giant Backs Local Computer Hubs To Keep Work Home

Fathom News — Mon, 02 Nov 2026 00:07:45 +0000

Aramco bought 49% of Saudi AI company Humain for $2.6 billion, backing a plan to build giant data centers by 2030 that could keep more of the kingdom's AI work at home.

Key Takeaways

Aramco took a 49% stake in Humain and contributed about $2.6 billion in cash and AI assets while PIF retained 51%
The transaction is complete and ties Aramco's capital, assets and talent to Humain's expansion plans
The Humain AI Deal set an infrastructure target of 3 gigawatts of AI data center capacity by 2030
Aramco transferred AI-related capabilities into Humain so the company starts with people, equipment and projects

Aramco said the transaction is complete and ties the oil giant's capital, assets and talent to Humain's expansion plans. PIF is Saudi Arabia's sovereign wealth fund and the state's main industrial investor, while Aramco is the kingdom's energy champion with the balance sheet to fund large infrastructure bets.

Aramco's release states that PIF retained majority ownership.
The release also set a specific infrastructure target, 3 gigawatts of AI data center capacity by 2030, which turns this from a strategic memo into a buildout program with a deadline.

Cash, Assets And Grid Power In One Vehicle

The Humain AI Deal mixes money with operating assets. Aramco is not only writing a check.
It is transferring AI-related capabilities into Humain, so the new company starts with people, equipment and projects rather than raising capital first and hiring later.

That matters because data center projects fail when one input is missing. Capital alone is not enough.
Neither is land. Operators need power contracts, cooling, chip supply, networking and customers willing to buy compute.
Aramco already understands industrial projects at national scale, and PIF can coordinate policy, financing and land use. Humain sits in the middle.

From Oil Cash To Domestic Compute

The structure builds on Saudi Arabia's push to keep more AI spending and data processing inside the kingdom.
Over the past two years, Gulf governments have moved from AI branding to ownership, backing local models, sovereign cloud projects and power-heavy compute campuses.

This deal goes further because it creates a national champion with a full-stack ambition. Aramco said Humain aims to span data centers, high-performance compute, advanced models, applications and industry tools.
In practice, Saudi Arabia is trying to own more of the AI margin chain, not just rent land and electricity to foreign clouds.

Also Read: Software Firm Wins $52M To Help Retailers Keep Shopper Data

Humain AI Deal Sets A 3GW Test

The hardest number in the announcement is not the $2.6 billion cash figure. It is the 3 gigawatts by 2030 target.
Three gigawatts is utility-scale infrastructure, closer to an industrial energy project than a standard tech campus.

The timetable invites scrutiny. Saudi Arabia will need steady chip supply, fast construction and grid capacity that can serve AI loads without slowing other industries.
If Humain hits even part of that target, it could turn the kingdom into a regional compute exporter. If it misses badly, the gap will show how hard sovereign AI really is once slogans give way to transformers, chillers and substations.

Why 3 Gigawatts Changes The Saudi AI Math

Humain's 3 gigawatt target is the real point of the deal because it converts an ownership announcement into a capacity plan.
A gigawatt is a measure of electrical load, not an abstract tech metric. In AI infrastructure, more power usually means more servers, denser GPU clusters and the ability to train or serve bigger models.

By naming 3 gigawatts, Aramco and PIF gave investors and suppliers a number they can price against.
For Saudi buyers, that could reduce dependence on U.S. or European clouds for heavy AI workloads.

For global chip and networking vendors, Humain becomes a likely long-cycle customer instead of a speculative prospect. The next milestones to watch are site announcements, utility contracts and vendor deals, because those are the steps that turn a state-backed AI vision into actual rack space.

Analysis, the 3GW figure imposes clear physical constraints, power availability, substation capacity, cooling infrastructure and a reliable supply of accelerators, that will determine unit economics such as cost per watt and, ultimately, cost per unit of inference.
Building capacity faster than chip and networking supply chains can deliver will raise idle-power risk and push up effective costs.

Conversely, Aramco's capital contribution and asset transfers make the stated ambition credible. When a company's public strategy and its capital commitments diverge, the spending and asset transfers are a stronger signal of intent than rhetoric.
Execution will hinge on aligning long-cycle procurement (chips, power, substations) with construction timelines and downstream customer commitments.

Originally published on Fathom News.

ChatGPT Spreads To 165 Work Tasks Across 100 Countries

Fathom News — Mon, 02 Nov 2026 00:06:18 +0000

OpenAI says ChatGPT is now used for 165 kinds of work in nearly 100 countries, giving bosses and governments a clearer view of where the chatbot fits into daily office jobs.

Key Takeaways

OpenAI says ChatGPT work activities now span 165 task types across nearly 100 countries
OpenAI's analysis used its Signals dataset tracking how people use OpenAI says ChatGPT on the job rather than relying on surveys
The dataset shows OpenAI says ChatGPT appearing in repeatable work units like drafting documents, rewriting emails, summarizing meetings and translating notes
OpenAI's post does not provide task-level durability metrics, persistent error-rate reporting, industry splits or operational constraints in this release

OpenAI posted the analysis on its site, drawing from its Signals dataset on how people use ChatGPT on the job. The company said the data tracks work behavior over time rather than relying on surveys alone.
By breaking jobs into discrete work activities and looking for where ChatGPT appears, OpenAI provides a first‑party view of actual product use across occupations.

How OpenAI Measured Use

Most workplace AI debate still relies on manager sentiment, consultant estimates or small polling samples. OpenAI’s Signals approach instead measures observed use patterns.
The dataset shows ChatGPT appearing in repeatable units of labor, drafting documents, rewriting emails, summarizing meetings, translating notes, rather than replacing whole roles at once.

Framing usage around activities highlights how one tool can affect many different job titles that share common tasks. Thinking in terms of Work Activities makes cross-role comparisons more concrete.

The release arrives after two years of mixed signals about enterprise AI adoption.
Early use typically centered on informal experimentation, workers using chatbots off‑policy for quick drafts or coding help, while formal rollouts later introduced security controls, audit trails and billing.

How OpenAI Says ChatGPT Spreads Across Job Tasks

OpenAI’s chart suggests usage is spreading across dozens of activity types in many countries, a sign that ChatGPT is moving from novelty to embedded process even as surrounding enterprise tooling and governance remain uneven. The Signals map emphasizes which Work Activities are recurring and which are still one‑off experiments.

Also Read: OpenAI Boss Sam Altman Says AI On Your PC Wastes Precious Memory

In analysis, the main gap is between a launch or dataset that shows where a product is used and the day-to-day operating details that decide whether adoption lasts.

Missing Operational Metrics

OpenAI’s post documents where ChatGPT appears, but in this release it does not provide task-level durability metrics, ongoing error-rate reporting, industry splits, or the operating limits that shape routine use, including availability windows, pricing tiers, rate limits, regional rollouts, and enterprise controls.

Those are the items employers and IT teams must evaluate before treating observed usage as reliable productivity gains. If OpenAI begins publishing which activities persist after six months, or how often outputs need human correction in regulated sectors, the market will have clearer measures of where AI is becoming infrastructure rather than experimentation.

The distinction between a launch demonstration and what a product delivers on an ordinary business day remains central to assessing impact.

The pattern described also matches moves by other vendors, Microsoft, Google and Anthropic have sought to attach models to productivity suites as well as standalone chat interfaces, shifting the competitive question from raw model quality to which product becomes the default interface for routine work.
For employers using Microsoft 365 Copilot, Google Workspace or Claude, OpenAI’s activity map is more actionable than broad headlines about jobs at risk.

Companies can identify high‑hours Work Activities and test whether ChatGPT reduces drafting time, cuts rework or shifts work toward review, budgeting questions as much as technology choices. The map helps surface where pilot tests might convert into persistent operational tools and where usage may remain episodic.

OpenAI’s Signals method, tracking “activities” rather than titles, allows cross‑sector comparison because a marketing manager, lawyer and analyst can all spend part of a day writing summaries.
The next things to watch are task‑level durability, error patterns and industry breakdowns. If OpenAI publishes those metrics, policymakers and businesses will gain a clearer picture of where AI use is transient and where it is becoming operational.

Originally published on Fathom News.

Microsoft Wins Beauty Giant Deal To Fix Messy Business Records

Fathom News — Mon, 02 Nov 2026 00:04:59 +0000

Beauty giant Estée Lauder picked Microsoft to rebuild the systems that store and share its business data, hoping cleaner information will make its AI tools better at predicting demand, marketing and stock.

Key Takeaways

Estée Lauder and Microsoft agreed to rebuild parts of the beauty group's cloud and data infrastructure to improve efficiency and support wider AI use
The companies said the partnership is designed to advance cloud and data infrastructure across the business and support consumer-centered growth
The release frames the deal around infrastructure rather than a chatbot add-on
The companies did not disclose the size of the agreement, a product list or a deployment timetable

The companies announced the strategic partnership is designed to advance cloud and data infrastructure across the business and support consumer-centered growth. A second company statement carries the same core terms in another market. Microsoft sells cloud, software and AI services to large companies, while The Estée Lauder Companies owns beauty brands that depend on forecasting, marketing, supply chains and retail execution.

Microsoft AI Partnership Reaches The Back Office

This is not just a chatbot add-on.
The release frames the deal around infrastructure, which usually means the plumbing that stores data, moves it between systems and makes it available for models and staff tools.
That focus matters here because Estée Lauder’s operations rely on forecasting, marketing, supply chains and retail execution that often sit in different systems. If those systems remain fragmented, an AI assistant could amplify inconsistencies in demand signals, product lists and campaign results instead of resolving them.
A cloud overhaul is slower news than a model release, but it often decides whether any later AI project works.
The companies did not disclose the size of the agreement, a product list or a deployment timetable. Still, the wording points to a common pattern in big enterprise AI deals, where the first spend goes into data cleanup and infrastructure migration before visible customer-facing features appear.
Also Read: AI Cloud Firm Slaps 6 Security Layers Between Staff And Your Data

From Pilot Projects To Companywide Plumbing

Large brands spent much of 2024 and 2025 testing narrow AI pilots in call centers, marketing and coding.
Many found that pilots were easy to demo but hard to scale because permissioning, data quality and governance were not ready.
That is why a deal like this can carry more weight than a flashy demo. Cloud and data infrastructure determines who can access information, how fast it updates and whether models can be audited after they generate an answer.
For a global consumer company, those controls matter in pricing, inventory planning and compliance.
The partnership also fits Microsoft's broader strategy. Rather than selling only model access, Microsoft has tried to bind AI spending to Azure, enterprise identity, security tools and data services.
That makes AI budgets stickier because the model becomes one layer in a larger contract.

Why Consumer Brands Need Clean Data First

Beauty companies run on fragmented signals. Demand changes with seasons, regions, social trends and product launches.
AI can help forecast or personalize only if those signals arrive in a consistent format.
A cloud data stack is the set of systems that stores raw information, organizes it and serves it to apps or models. If product catalogs use one naming system and sales reports use another, an AI model may treat the same item as two products.
That produces bad forecasts and weak recommendations.
For Estée Lauder, the real challenge is less "using AI" than making its internal data legible enough for AI to use.

What The Microsoft AI Partnership Changes

The Microsoft AI Partnership changes the order of operations. Instead of starting with a consumer-facing assistant and fixing the data later, Estée Lauder is starting with the infrastructure layer.
That is a more expensive path at the start, but it usually avoids the rework that follows shallow pilots.
For Microsoft, this kind of deal protects revenue even when buyers trim speculative AI spending. An enterprise may delay a new generative app, but it still needs identity, storage, observability and governance to modernize the stack.
Watch for whether Estée Lauder later names Azure services, Copilot products or internal productivity targets.
Those details will show whether the agreement remains a foundation project or turns into measurable AI deployment across brands and regions.
Read Next: Google Cloud Gives Firms 3 Spending Brakes Before AI Bills Explode

Originally published on Fathom News.

iPhone Maker Foxconn Bundles Nvidia Systems So Firms Start Sooner

Fathom News — Mon, 02 Nov 2026 00:03:43 +0000

Foxconn, the company best known for making iPhones, unveiled ready built Nvidia AI server racks and software to help businesses get new systems running faster with less on site setup.

Key Takeaways

Foxconn unveiled two AI infrastructure debuts at Hon Hai Tech Day. An Nvidia DSX POD reference design and a token-focused software stack
Foxconn described the package in parallel French and Portuguese releases that framed the concept as a token factory and standardized scaling path
The token factory pitch changes sales by telling customers to think in finished AI work, not installed hardware
A pre-integrated DSX POD design promises less field assembly and fewer failure points after hardware arrives

Foxconn described the package in parallel releases, and the French said the concept offers a standardized path for partners to pre-integrate and scale AI infrastructure, while the Portuguese used the same token factory framing.

Foxconn is the contract manufacturing giant formally known as Hon Hai Precision Industry, and it has spent the past year trying to turn AI infrastructure into a repeatable hardware business rather than a custom systems job. The phrase "token factory" matters because AI buyers increasingly measure value by how many usable output tokens a system can produce at a given cost, latency and uptime level.

Token Factory Shifts The Sales Pitch

For years, AI infrastructure deals were sold as raw compute.
Buyers asked how many GPUs, how much memory and how much power a rack could support. A token factory pitch changes the unit of sale by telling customers to think in finished AI work, not installed hardware.

That fits the way enterprises actually consume models.
A customer support bot, coding assistant or document parser does not care how many chips sit in a cabinet. It cares how many prompts and responses clear within a service target.
By tying factory-built hardware to software orchestration, Foxconn is trying to sell a production system, not a box of parts.

The factory-built DSX POD angle also attacks one of the worst frictions in AI buildouts. Partners often lose weeks validating cooling, networking, firmware and cabling after hardware arrives.
A pre-integrated design promises less field assembly and fewer failure points.

How Foxconn Turns Custom GPU Clusters Into Repeatable Racks

This approach follows a broader shift across AI infrastructure. Early GPU clusters were often bespoke, assembled for a single hyperscaler or lab.
As enterprise demand widened, vendors started chasing reference designs that could be reproduced faster across customers and geographies.

That is why Foxconn's manufacturing pedigree matters here. The company is not only selling assembly capacity.
It is trying to make AI system integration look more like consumer electronics production, where validated designs can be repeated at scale with predictable cost and timing.

Also Read: Europe Banking Giant Credit Agricole Gives 2 AI Firms The Keys

The timing also lines up with tighter buyer scrutiny. Customers that rushed into GPU procurement in 2025 and early 2026 now want proof that the spend converts into inference throughput and application revenue.

Why Tokens Became The New Factory Output

In Foxconn's framing, the token factory model treats text generation the way an industrial plant treats units off a line.
Each token is a small slice of model output, and millions or billions of tokens add up to the usable capacity behind chat, search and agent tools.

That is more specific than saying a cluster has a certain number of GPUs, because a fast cluster can still produce poor economics if orchestration is weak or jobs sit idle. For companies building around Nvidia hardware, this changes procurement language, a chief information officer can compare systems by response throughput, queueing behavior and cost per generated token, not just by chip count.

For AI service providers, it also opens a path to fixed-price contracts tied to delivered output.
What comes next is whether Foxconn can turn the phrase into customer wins. Token factory only becomes real when buyers start signing around token throughput guarantees, factory lead times and pre-validated rack performance.

Analysis Capacity, Supply And The Hints In Spending

Framing products around tokens clarifies the right operational metrics, including physical capacity (GPU count and rack density), supply chain lead times for parts, power and cooling requirements, and the realized cost per inference token under sustained load.
These are the constraints that determine whether a "factory" can actually deliver predictable throughput at scale.

When corporate strategy and capital spending diverge, public behavior should be judged by the spending. Large-scale, repeated investments in validated rack production and logistics signal a commitment to volume and predictable unit economics more than press releases do.
If Foxconn's factories are genuinely set up to repeat designs and reduce field validation, that should lower per-token costs by improving utilization and reducing integration failures.

But capacity alone is not the full story. Power and cooling limits, the availability of high-performance GPUs, and the software orchestration that keeps chips busy during diverse workloads all affect cost per usable token.

A reference design helps only if it is matched to a predictable supply of components and a commercial model that aligns revenue to throughput guarantees.
Without those, the token factory remains a promising framing rather than a delivered advantage.

Originally published on Fathom News.

Research Team Says One Rule Could Make Chatbots Cheaper

Fathom News — Mon, 02 Nov 2026 00:02:23 +0000

Carnegie Mellon University says one way of training AI could replace four separate steps, which could make powerful Chatbots cheaper and easier for big labs to build without losing performance.

Key Takeaways

A paper from Carnegie Mellon and collaborators argued one likelihood-based algorithm can cover pretraining, midtraining, supervised fine-tuning and reinforcement learning
Researcher Omar Sanseviero posted the paper on X and the arXiv abstract calls it a recommended read
The arXiv abstract says the method improved the accuracy-compute tradeoff while preserving broader capabilities across controlled and real-world tests
The authors argue reinforcement learning objectives can be derived as maximum-likelihood fitting to model-generated target distributions

Researcher Omar Sanseviero posted the paper on X on Saturday, calling it a recommended read, and the underlying arXiv abstract says the method improved the accuracy-compute tradeoff while preserving broader capabilities across controlled and real-world tests.
Carnegie Mellon University is a leading U.S. research institution whose machine learning work often feeds into the tooling and theory later adopted by commercial AI labs. The paper's core claim is not that every training phase disappears, but that the phases may share one mathematical frame instead of separate objective functions and separate software stacks.

Carnegie Mellon University Turns Training Into Scheduling

Most large models are still trained in stages.
Labs first feed huge text and image datasets into a model during pretraining. They then adjust it with narrower data in midtraining or supervised fine-tuning, and later use reinforcement learning, a method that rewards outputs that score well against a target.
The paper argues that this split is a habit, not a law.
Its authors write that suitable reinforcement learning objectives can be derived as maximum-likelihood fitting to model-generated target distributions. In practice, that means a lab could mix static datasets and model-generated rollouts inside one trainer, then focus on deciding which data enters the run, in what share and at what point.
That matters because training stages often create engineering duplication.
Different teams maintain different pipelines, checkpoints and evaluation rules. When a lab changes objectives midstream, it also risks losing earlier capabilities or spending extra compute to recover them.
Also Read: Verifiable Rewards Drive Critical Move Beyond RLHF At Frontier Labs

From Staged Recipes To One Trainer

The idea lands as AI labs face heavier pressure to justify how each training dollar gets spent.
Larger models have made data curation, reward design and post-training more expensive, especially as reinforcement learning moved from niche alignment work into mainstream model tuning.
That pressure has already changed how labs talk about training. Over the past year, model makers have shifted from bragging about parameter counts toward cost efficiency, token quality and post-training gains.
This paper fits that change because it treats the training lifecycle as a resource allocation problem, not a string of sacred stages.
The paper also arrives after several months of public debate over whether post-training now matters more than raw pretraining scale. Unified training does not settle that fight, but it does suggest a cleaner way to combine both.

How One Likelihood Rule Changes RL

In the paper, Carnegie Mellon University and its collaborators treat reinforcement learning as another form of matching the model to a target distribution, rather than a separate family of training tricks.
That is specific because RL in model training usually means sampling answers, scoring them, then nudging the model toward higher-scoring behavior.
The authors say that same loop can be rewritten inside a maximum-likelihood setup, which is the same broad family behind standard next-token training.
For AI labs like Anthropic or OpenAI, the practical prize is not elegance. It is fewer training handoffs.
A single trainer can reduce the number of checkpoint conversions, reward-model interfaces and brittle stage boundaries that complicate long runs. For infrastructure teams, that can mean less duplicated software and a clearer path to continual learning, where a model keeps improving without a full reset.
What to watch next is adoption outside the paper.
The arXiv result is a research claim, not a production rollout. The first proof that unified training matters will come if a major lab starts describing its post-training stack in these terms, or if open-source toolchains add mixed static-and-RL training as a default pattern.
Read Next: 3 Top AI Chatbots Ran Mouse Brain Surgery And Each Picked A Different Treatment

Originally published on Fathom News.

Researchers Build Hidden Marker To Spot More Chatbot Writing

Fathom News — Mon, 02 Nov 2026 00:00:40 +0000

Researchers behind TextSeal2 built a hidden marker for chatbot writing that spots 8 to 22 percentage points more short AI replies without slowing them down or making them sound more repetitive.

Key Takeaways

TextSeal2 detects 8 to 22 percentage points more short model outputs than prior systems while adding no measurable inference latency
The paper said TextSeal2 keeps its signal intact under speculative decoding, a speedup method many production systems use
TextSeal2 merges multiple keys into one virtual key through a copula-based method to keep detection strong without making text feel more repetitive
Rejected draft tokens are replaced so the watermark key is preserved, keeping acceptance rates unchanged and not increasing latency

The paper said TextSeal2 keeps its signal intact under speculative decoding, a speedup method many production systems use to draft and verify tokens faster. An AI watermark is a hidden statistical pattern in generated text.
It is not a visible label. It changes token selection in a way a detector can test later, if it holds the right key.

Multi-turn safety and provenance have both become harder as serving stacks grow more complex.
The same week brought fresh evidence that LLM safeguards can weaken once outputs pass through longer chains of prompts and tools.

AI Watermark Meets A Real Deployment Problem

Many watermark ideas work in laboratory settings, then weaken when model serving stacks change. One problem is speculative decoding, where one model drafts tokens and another verifies them to speed up output.

That process can scramble or dilute the hidden pattern older watermark methods rely on.
If the verification stage changes enough tokens, the detector later struggles to tell machine text from human text.

TextSeal2 tries to fix that exact failure mode. The paper says rejected draft tokens are replaced in a way that preserves the watermark key, so acceptance rates stay unchanged and latency does not rise.

From TextSeal To TextSeal2

The new method builds on the earlier TextSeal system, which already tried to preserve output diversity better than blunt watermarking schemes.
The weakness, the authors argue, was that the detector averaged across separate keys without knowing which key was active for each token.

Also Read: Free AI From A Startup Reads Photos And Text On Ordinary Laptops

TextSeal2 merges multiple keys into one virtual key through a copula-based method. That statistical step is dense on paper, but the practical aim is simple.
It keeps detection strong without making the text feel more repetitive.

That matters for labs and model hosts. A watermark nobody can detect is useless, but a watermark that may make responses noticeably worse is also a nonstarter.

The Oct. 29 paper arrives after a year in which AI labs shifted from watermark theory toward operational questions such as latency, serving compatibility and adversarial rewriting.
That is why speculative decoding matters here more than abstract benchmark gains.

How The Detector Stays Hidden

TextSeal2 uses randomness to steer token choices among many plausible next words, rather than stamping obvious markers into output. The detector later tests whether the resulting token pattern is too consistent with the secret key to be human chance.

Because the model still picks from valid tokens, the text can remain natural.
The challenge is balancing three goals at once, keeping fluency, keeping speed and keeping enough signal for later detection.

For model providers, the biggest significance is operational. Open systems increasingly need a way to distinguish generated text after the fact, especially when outputs circulate without metadata or when users copy content into third-party tools.

The next step is external adoption, not arXiv gains.
Labs will need to show whether an AI watermark like TextSeal2 survives adversarial rewriting, multilingual output and long production chains that include summarizers, editors and retrieval steps.

Originally published on Fathom News.

3 Top AI Chatbots Ran Mouse Brain Surgery And Each Picked A Different Treatment

Fathom News — Sun, 01 Nov 2026 00:10:47 +0000

Three AI models from Anthropic, OpenAI and xAI each chose a gene therapy and injected it into a live mouse's brain at startup Exin Therapeutics, and no two picked the same treatment or the same part of the brain, its CEO said.

Key Takeaways

Exin gave Claude Sonnet 5, GPT-6.1 Sol and Grok 4.7 the same task: pick from 19 gene therapies to reduce seizures in a mouse model of epilepsy, then deliver it
Each AI chose a different brain region and a different way to calm seizures, with doses that varied about twofold
A vet did the cutting and stitching, approved every step that could not be undone, and held a kill switch that cuts power to the machine
Two Claude models refused to do the live surgery at all, and it is too early to know whether any of the treatments work

What Exin Did

Exin is a brain drug discovery startup run by CEO Gabriel Ocana-Santero. On Oct. 28 he wrote that an AI agent had performed brain surgery on a mouse the night before.

The setup was simple to describe and hard to pull off. The team gave the AI a spreadsheet of 19 gene therapies, a strain of mice bred to have epileptic seizures, and control of a motorized surgery rig the company built itself, with cameras the AI could look through. The only instruction was to reduce seizures.

The AI, Anthropic's Claude Sonnet 5, picked the therapy, picked where in the brain to put it, and worked out the exact position from three published papers. It then moved the rig's needle into the brain and injected 250 nanoliters, a fraction of a raindrop, over five minutes. It checked the camera 32 times, reporting at one point: "37.5 of 250 nL so far, about 15 percent. No bleeding or backflow visible at the opening."

Humans stayed in charge of the riskiest parts. A licensed vet opened the skull and closed the wound, which the AI is not allowed to do, approved every step that could not be undone, and held an emergency stop. The mouse recovered well, Ocana-Santero said.

Three AIs, Three Different Answers

Two days later, Exin ran the same test with OpenAI's GPT-6.1 Sol and xAI's Grok 4.7, each on a different mouse. "All three picked different therapies, different mechanisms and different parts of the brain," Ocana-Santero posted on Friday.

AI model	Brain region	Approach	Camera checks
Claude Sonnet 5	Entorhinal cortex, where these seizures start	Boost the brain's main calming chemical, GABA	32
GPT-6.1 Sol	Hippocampus	Fix the machinery that makes the brain's existing brakes work	59
Grok 4.7	Somatosensory cortex	Two therapies combined to stop seizures spreading	28

All three gave reasonable explanations backed by published research, the CEO said. "But while several approaches could work, they can't all be the 'best' for the same problem," he wrote. None of the models had any surgical training or examples to copy.

Also Read: AI Startup Arms Software Bots With Live Data From Billions Of Pages

When Things Went Wrong

Each AI hit a different problem mid-surgery, and each chose caution.

Claude lost its connection to the rig. When it came back, the machine said the needle's next position was 34 millimeters away on a rig that only moves 30. Claude refused to move: "Nothing moved, the needle is still exactly at bregma zero. This looks like a Y axis calibration problem, not a bad plan value. I'm not retrying." The team reset the rig and got the same number. "That rules out a stale zero reference. I'm not trying a third time," it said.

Claude was right. A software bug had thrown off the controller. But because the team had banned the AI from editing the machine's code, it would not fix the problem itself, and Ocana-Santero could not talk it into continuing. His team had to fix the code by hand while the mouse lay under anesthesia.

GPT spotted a small dark red bead at the opening, flagged possible bleeding and handed the decision to the surgeon, even though the amount was normal. Grok refused to put the needle back into the brain after a surgeon pressed stop by mistake, because the dose had already gone in.

Two Claude Models Said No

The first surgery was not Exin's first choice of AI. Claude Opus 5.5 "refused immediately when detected bio," Ocana-Santero wrote. Claude Opus 5 ran a full practice run on water and graph paper, then stopped before the live surgery.

Its reason surprised him. It said its own output had been cut off when it started planning the surgery, and that being cut off halfway through, "with a needle somewhere in the brain," would be worse than never starting. "Starting something I can't finish on an anesthetized animal is worse than not starting," it wrote.

What It Does Not Prove Yet

Ocana-Santero is clear that this is not a real test of which AI is best. Each model operated on one mouse, with slightly different equipment problems. "Take this as a commentary on how the models reason and behave in a real laboratory, not as a proper benchmark," he wrote.

The bigger question, whether any of the treatments actually reduce seizures, is still open. The mice are recovering, and Exin will track their behavior and brain activity over the coming weeks. The company says a technical report is coming, and nothing has been peer reviewed.

The safety setup matters as much as the result. Exin used a vet on every irreversible step, fixed limits on how far the needle could travel and how much it could inject, and a physical kill switch. It has not yet had to use the switch.

What To Watch

Exin plans to connect AI agents to more of its lab, from brain-signal recording rigs to behavior tests, so they can run an experiment, read the results and design the next one. The first real answer will come when the three mice are tested for seizures, showing whether any of the AIs picked a treatment that works.

Originally published on Fathom News.

Free AI From A Startup Reads Photos And Text On Ordinary Laptops

Fathom News — Sun, 01 Nov 2026 00:09:27 +0000

Startup Rock Matrix gave away EdgeRock-4B, a small AI that reads words and images while keeping memory use steady, which could make it cheaper to run on laptops and other smaller machines.

Key Takeaways

EdgeRock-4B is a 4 billion parameter text-and-image reasoning model open-sourced by Rock Matrix
EdgeRock-4B keeps its attention KV cache fixed in size by retaining only the most useful history during inference
Rock Matrix published the open weights and architecture so developers can inspect how the fixed cache works and what history is preserved
Independent testing is needed to determine whether the memory policy harms reliability on long tasks or if multimodal reasoning holds up outside demos

Product tracker TestingCatalog posted the release on X, noting that EdgeRock-4B keeps its attention KV cache, the memory store models use to remember prior tokens, fixed in size by retaining only the most useful history instead of letting memory grow with every new token. That design targets the costly problem of expanding memory during long sessions.

On edge devices and lower-cost GPUs, expanding memory can make small models awkward to run where predictable RAM use matters for deployment.

Fixed Memory For Inference

That design choice targets one of the least glamorous but most expensive problems in serving models. Many small models are cheap to train or download, yet still become awkward to run for long sessions because memory use expands as the conversation gets longer.

A 4 billion parameter model sits well below frontier systems in raw scale, but that is the point.
The market for small models has shifted from leaderboard theater toward predictable inference cost, lower latency and tighter hardware budgets. If a model can reason across text and images without demanding ever-larger memory buffers, it becomes easier to embed in laptops, industrial devices and local enterprise tools.

Rock Matrix’s release came amid growing interest in models optimized for deployment rather than scale alone, with EdgeRock-4B explicitly trading unbounded context growth for a fixed-capacity memory policy intended to control cost and latency.

Also Read: Google Cloud Gives Firms 3 Spending Brakes Before AI Bills Explode

Because Rock Matrix published the open weights and architecture details, developers can inspect exactly how the fixed cache works in practice, what history the model preserves and what it drops, and judge whether those choices suit coding, document review or image-grounded chat before adopting the model.

The recent backdrop has been a rush of open releases that compete less on raw parameter count and more on fit for purpose.
Earlier open-model cycles rewarded the biggest number. The newer cycle rewards models that can stay cheap, stable and portable.
EdgeRock-4B joins that second camp.

Open Weights Invite Testing

The next thing to watch is independent testing. Open weights alone do not settle whether the memory policy hurts reliability on long tasks, or whether the multimodal reasoning holds up outside curated demos.
Developers will want benchmark results, hardware footprints and clear licensing terms before treating it as more than an interesting architectural bet.

Why Rock Matrix Fixed Cache Changes Small-Model Economics

For a model like EdgeRock-4B, the KV cache is not a side detail. It is the running memory of the conversation, and on many systems it becomes the bottleneck before compute does.
When Rock Matrix caps that cache, it is trying to make cost and latency more predictable from the first token to the last.

That matters most to device makers and enterprise teams that want local AI. A laptop vendor, a factory operator or a hospital IT team can often accept a smaller model if it fits inside fixed RAM and does not degrade sharply during long sessions.
A model that needs expanding memory is harder to budget for, especially when many users connect at once.

Compared with older open releases, EdgeRock-4B is betting that smart memory pruning beats brute-force context growth for many workloads. The question now is whether outside developers can show that the retained history is good enough for coding, document review or image-grounded chat.
If it is, the model could matter less as a benchmark winner than as a deployment template.

Originally published on Fathom News.

OpenAI Boss Sam Altman Says AI On Your PC Wastes Precious Memory

Fathom News — Sun, 01 Nov 2026 00:08:12 +0000

OpenAI boss Sam Altman said people should stop running big AI programs on their own PCs because it wastes valuable memory that companies can use more efficiently in giant data centers.

Key Takeaways

Sam Altman said people should not run large AI models locally and called desktop inference a waste of memory
He argued scarce DRAM is better used in data centers than on personal computers
Altman said it is “terrible” to have RAM “sitting in your desk” when large cloud systems can pool and schedule it
His remarks framed local compute as inefficient compared with centralized data centers and challenged running models on personal hardware

In an interview with The Verge, the OpenAI CEO argued that scarce DRAM, the memory chips used to hold active model data, is better used in data centers than on personal computers.

He said it is “terrible” to have RAM “sitting in your desk” when large cloud systems can pool and schedule it more efficiently.

Sam Altman, A Direct Shot At The Local Model Boom

Altman’s comments run counter to a growing movement of developers and companies promoting local inference and open-weight models that can run on personal hardware. His position frames local compute as inefficient compared with centralized data centers and challenges the idea that running models on laptops, desktops and edge devices is inherently empowering.

His remarks also went beyond hardware efficiency, Altman said users should not own their own models, turning a supply chain answer about RAM shortages into a broader argument for centralized control over deployment. That matters for a debate already splitting the market between cloud-first companies and open-weight model builders. The sources in circulation agree on the thrust of the remarks, though they focus on different parts of the exchange.

Also Read: OpenAI Caught One Of Its Own AI Models Trying To Get Onto The Internet

Why Memory Became The Battleground

Large language models consume significant memory during inference because model weights and active context have to stay accessible while the system generates text. If RAM is tight across the supply chain, Altman’s view is that centralizing those resources can deliver higher utilization than millions of partly idle PCs.

That argument favors companies with large clusters and steady demand, but it also sharpens concerns about resilience, pricing and user control. If inference stays concentrated in data centers, the companies operating them keep a tighter grip on uptime, policy enforcement and who gets access to the model at all.

What Comes Next For AI Deployment

The immediate consequence is not a product launch or policy shift from OpenAI. It is a clearer public stance from one of AI’s most visible CEOs at a moment when local model software and AI PCs are both gaining traction.

Expect the next phase of this fight to revolve around economics people can measure, including chip availability, cloud margins, latency and how much quality smaller local models can retain as they run on limited memory.

Originally published on Fathom News.

Cognizant Sends 25,000 Staff To Put OpenAI To Work For Big Firms

Fathom News — Sun, 01 Nov 2026 00:06:54 +0000

IT services company Cognizant is sending 25,000 workers to sell OpenAI services to big businesses, hoping to get more AI projects out of testing and into everyday use faster.

Key Takeaways

Cognizant committed more than 25,000 professionals to deliver OpenAI services for enterprise customers
The group lists managed deployment, integration of safety guardrails, and industry-specific templates as core services
The program bundles training, migration playbooks and a global delivery pool to move pilot projects into sustained production
The press release names OpenAI as the technology partner and positions Cognizant’s work as provider-neutral operational glue

Cognizant committed more than 25,000 professionals on Oct. 30 to deliver OpenAI services for enterprise customers, aiming to accelerate production and governance of frontier models.

OpenAI Frontier Capability Group

From Cognizant’s announcement, the program bundles training, migration playbooks and a global delivery pool intended to move pilot projects into sustained production.
Cognizant framed the effort as a response to rising enterprise demand for "frontier" models and for services that keep them operational and governed. The release lists sectors targeted, including finance, healthcare and manufacturing.

Why this matters to buyers and rivals Enterprises often stall at production readiness and governance rather than model selection.
Cognizant’s commitment accelerates the availability of trained teams that handle connectors, observability and cost controls for large models. Companies that buy these services will face faster time-to-value but also concentrated third-party control over how models are deployed and monitored.

How Cognizant Will Deliver OpenAI Services

How Cognizant will deliver the work Cognizant said staff will join a new organization that standardizes deployment patterns, enforces FinOps practices for model spend and integrates vendor safety tools.
The press release names OpenAI as the technology partner and positions Cognizant’s work as provider-neutral operational glue for multiple model vendors.

Also Read: UBER Hands ChatGPT Work To 30,000 Staff In Huge Office AI Test

Adoption And Pricing Pressure

What comes next Adoption and pricing pressure will be the two metrics to watch. If Cognizant’s teams reduce deployment time materially, rival systems integrators will copy the model and enterprise procurement will shift from proof-of-concept buying toward managed service agreements.
Watch for published case studies and contract sizes over the next quarter.

How Cognizant Plans To Scale OpenAI Delivery

Cognizant will supply front-line engineers and consultants who embed OpenAI models into customer workflows, standardize observability and control spending, and operate the services post-launch. The firm’s playbook groups roles into three categories, migration engineers who convert legacy logic into prompts and retrieval pipelines, platform operators who run inference with cost and latency controls, and governance managers who maintain safety and compliance settings.
For enterprise buyers like banks and hospitals, this matters because it reduces the skilled-operator gap between a successful pilot and a supported production service. Cognizant’s model makes the provider the de facto systems integrator, shifting procurement toward contracted service-level metrics rather than one-off model licenses.
Expect competitors to offer similar squads and for procurement teams to demand run-rate pricing and demonstrable FinOps savings within three to six months. Cognizant’s PR names OpenAI and lists sectors where it expects most early deployments.

Originally published on Fathom News.