AI News Today, Nov 10: Anthropic Releases Fable and Mythos 5.2
Anthropic releases Fable and Mythos 5.2 with shared capability limits; Amazon changes its ticket classifier; Notion reports exposed AI chat transcripts; Google's Gemini review policy and new TPU systems; and LingBot-Touch research for robot touch. This audio was created with Boson AI's Higgs Audio -- https://www.boson.ai/higgs-audio
Listen on Spotify · Watch on YouTube · Pocket Casts
Anthropic releases Fable and Mythos 5.2
Eris: Anthropic has released Fable 5.2 and Mythos 5.2, two versions of the same underlying AI model with different access rules and safeguards. <|prosody:expressive_high|>The company says both also share a training intervention that holds back certain skills used to accelerate AI development.
Vestra: Fable is broadly available, with stronger safeguards around high-risk cyber and biology requests. Mythos is offered to verified organizations with certain of those safeguards removed. The shared training change is separate: Anthropic says it limited improvements in accelerator-kernel optimization and language-model architecture research to roughly the level of its earlier Opus 5.5 model.
Eris: An accelerator kernel is a small piece of code that helps run calculations efficiently on AI hardware. That makes the boundary consequential for people trying to speed up smaller local models too, not just companies building frontier systems. <|prosody:expressive_high|>That's the spillover I'd worry about. Anthropic says agentic coding and conventional machine-learning research were unaffected, but it hasn't established that adjacent work has no spillover. Would choosing Mythos change that shared limit?
Vestra: No, and this is the distinction that matters. A request filter can block an answer at the point of use; the training intervention changes the capabilities both versions share. Anthropic also describes a separate, narrow serving-time fallback for AI research topics. <|prosody:expressive_high|>Mythos removing certain safeguards does not remove <|prosody:pause|>the shared training limit.
Eris: The release also comes with a strong result on a test of adapting to unfamiliar interactive puzzles. <|emotion:surprise|>ARC Prize reported a much higher result with Anthropic's own adapter than with the standard setup. That adapter is software that preserves the model's reasoning between calls and compresses long conversations, so the surrounding software is part of the result. <|prosody:expressive_high|>That's what catches my attention: how much the setup matters here.
Vestra: And I'd keep the cost in view alongside that result. The cost of a completed task can differ from the speed of generating individual words. Artificial Analysis ranked Fable highly in its own evaluation but reported lengthy, costly tasks in that particular run. <|prosody:expressive_high|>That's a property of that demanding test, not a price for every request.
Eris: <|emotion:surprise|>Anthropic also reports using a thousand AI agents to formalize an existing proof of Fermat's Last Theorem in Lean, a system that checks mathematical proofs. <|prosody:expressive_high|>I'm impressed by the formalization, but they didn't discover the proof. The company's linked physics work makes narrower claims than solving black-hole physics, and independent replication hasn't been established.
Amazon switches its ticket classifier
Eris: Amazon says it replaced a general-purpose AI model in an internal ticket-classification system with a specialized model called Strands Decider. Its engineers report lower modeled serving costs and say the replacement performed better on a separate ticket test, though the details needed to independently judge that test aren't public.
Vestra: The job is to choose among twenty ticket categories. Claude Sonnet had generated a structured answer; Decider instead scores the available choices directly using a decision component attached to a pretrained model. <|emotion:enthusiasm|>That's the fit I like here: for a fixed set of labels, it doesn't need to write out the answer token by token.
Eris: <|emotion:surprise|>That makes the cost comparison appealing. The case study's list-price calculation puts Decider at roughly five hundred dollars per million tickets, against about twenty-two thousand for Sonnet, under its stated usage assumptions. <|prosody:expressive_high|>But those are <|prosody:pause|>modeled serving costs, not verified annual savings. The report doesn't publish actual utilization, ticket volume or invoices.
Vestra: The quality claim has similar limits. The engineers say Decider beat Sonnet, but don't give the test size, the correct labels or the score. Before the switch they ran both systems in parallel while applying Sonnet's answers.
Eris: But if they were applying Sonnet's answers, how would they know when Decider was right to disagree?
Vestra: <|sfx:humming|>Hmm, that's what I'd want the independent reference for. Running them in parallel finds disagreements; it doesn't tell you which answer was right without that reference.
Eris: And the two systems weren't given precisely the same input. Long tickets lost their oldest comments to fit the new model's input window, and category descriptions were compressed. <|emotion:surprise|>The team also found that changing the order of fields in the structured input could shift its confidence, so they sorted those fields consistently. That sensitivity bothers me more than the tidy cost comparison reassures me.
Vestra: I'd give the fix some credit: that addresses the ordering issue they found, though it doesn't establish robustness to every change in a ticket. A conventional classifier trained on labeled examples is another possible comparison. <|prosody:expressive_high|>This case shows why a narrow decision model is worth testing, while leaving the actual bill and error rates open.
Notion chat transcripts exposed in Fireworks incident
Eris: Notion says a security incident at its AI provider Fireworks affected a dataset containing customer AI chat transcripts used for Notion Insights classification. Notion says fewer than one percent of workspaces may be affected and advises affected customers to treat the material as <|prosody:pause|>exposed.
Vestra: Notion says Fireworks confirmed the dataset on November eighth. The company says the vast majority of affected transcripts lacked user and workspace labels and the results of tool calls. <|prosody:expressive_high|>But the conversation text itself can contain identifying details or pasted credentials. That's why the missing labels don't reassure me. Notion recommends inspecting the exposed data and rotating credentials that appeared in chats.
Eris: Notion says it disconnected from Fireworks and deleted its data from the provider's system by November third, before the later confirmation of this customer dataset. It also disabled Insights. <|sfx:sigh|>Uh, I'd like that deletion to be reassuring. But deleting the provider-side copy doesn't establish what happened to any copy an attacker may hold.
Vestra: Fireworks describes two October activity windows: unauthorized access to environment variables in evaluation jobs, then a separate third-party provider incident followed by access to Fireworks development machines and an unauthorized workload. Its expanded disclosure includes some secrets, datasets and incomplete model artifacts, but doesn't assign each category to a particular event. <|prosody:pause|>We don't have a complete public reconstruction of the route to the Notion data.
Eris: Fireworks also says its Inference, Training SDK and Nexus products retain no customer data by default and that it found no impact to those products. How does that fit with the exposed Insights dataset?
Vestra: That default can coexist with an exposed Insights dataset. Neither company has explained the exact transcript storage path or which retention terms applied to it. <|prosody:expressive_high|>That's the missing detail I'd want before treating the default as reassurance here.
Gemini connected-app material and human review
Eris: Google's consumer Gemini policy says human reviewers may see a subset of interactions and material derived from connected Gmail and Drive content. <|emotion:surprise|>Google says a summary can sometimes be the entire relevant email or file, so connecting an app has consequences beyond what appears in a chat answer. I'd have expected a summary to contain less than the whole file.
Vestra: That possibility concerns me too, but the scope matters. Connected apps let Gemini retrieve relevant information to respond to a request. Google says selected interactions and derived material can be reviewed for quality, safety and improvement. <|prosody:expressive_high|>It does not say every connected email or file is reviewed, and it says it doesn't train directly on the entire inbox or Drive.
Eris: A public report alleges that casual prompts can trigger retrieval without explicitly naming the app. Its author asked Gemini to print a supposed hidden payload. <|prosody:expressive_high|>I'd want to see the actual trace: that's a model-generated answer, not a trace of what the backend actually retrieved or sent for review. The exact payload and review selection remain unverified. Would disconnecting the app erase what was already saved?
Vestra: Not necessarily. Disconnecting an app removes a future connection, but doesn't necessarily erase information already saved in Gemini activity. The controls do different things.
Eris: What changes if Keep Activity is off?
Vestra: Turning off Keep Activity changes how future chats are used and makes Workspace unavailable as a connected app on this consumer path. Google still describes limited retention and possible review for protection or feedback, and says previously reviewed chats and related data may be kept for <|prosody:pause|>up to three years.
Eris: <|prosody:expressive_high|>That retention is the part I'd want to check carefully, and account type matters here. Qualifying managed Workspace services have separate protections, so the consumer policy shouldn't be assumed to cover every enterprise deployment. Check the terms for the product and account you're actually using.
Google announces TPU 8tx and 8i Max systems
Eris: Google has announced two AI accelerator system configurations: TPU 8tx, aimed at confidential workloads, and TPU 8i Max, which packages two 8i chips together. The announcement describes different ways to build a system around existing chip designs.
Vestra: For 8tx, Google describes the 8t silicon foundation with AMD host processors and confidential-computing features intended to protect data and model weights while they're being processed. <|prosody:expressive_high|>That protection interests me, but whether it fits a customer's needs depends on details such as who can access the system and how the hardware verifies its identity. The announcement doesn't provide an independent security assessment or a detailed threat model.
Eris: For 8i Max, Google says the two chips in one package double compute, on-chip memory and high-bandwidth memory bandwidth compared with 8i. <|emotion:enthusiasm|>I can see why that packaging is appealing. More bandwidth can help move data to a model faster, but that doesn't automatically mean twice the useful work gets done. What else determines how much of that promise you get?
Vestra: The workload, software and communication between components all matter. Google hasn't supplied public prices, a delivery date or workload measurements that establish performance per dollar or watt for these systems. <|prosody:expressive_high|>I'm interested in the packaging too; it could matter substantially. We need measurements on the actual configurations to know how much.
LingBot-Touch learns from different touch sensors
Eris: Researchers have introduced LingBot-Touch, a model that turns readings from different kinds of touch sensors into signals robot controllers can use. In their experiments, the authors report that adding those learned signals improved average success across four controllers on a set of real-world manipulation tasks.
Vestra: Touch helps when a camera can't see what's happening under a finger: whether a grip is slipping, for example. <|prosody:expressive_high|>But sensors produce very different readings. Some respond to changes in electrical resistance, some to magnetic fields, and others use a camera to observe a deforming surface.
Eris: The researchers collected touch data from sensor-equipped gloves and remotely operated robot hands. Their model maps sensor locations into a common fingertip coordinate system, then uses sensor-specific input components and a shared network to process the signals. The model supplies features to a controller; it doesn't choose the robot's actions by itself. <|prosody:expressive_high|>What does it learn to predict when those readings are so different?
Vestra: During training, the system hides some sensor readings and learns to predict their internal representations from the surrounding touch information. It isn't simply copying raw readings. <|emotion:enthusiasm|>That's the part I find exciting: much of a touch recording can be inactive background, while the contact itself is sparse. That makes this useful to investigate.
Eris: For the real-robot comparison, the authors tested twenty-six tasks across two hand platforms, including sorting, opening a zipper and moving delicate objects. They compared the original controllers, controllers given raw touch, and controllers given LingBot-Touch features. <|emotion:enthusiasm|>On average, the learned features improved success more than raw touch for each of the four controller families, though they weren't best on every individual task. I like seeing that comparison with raw touch: it gets at what the learned features contribute.
Vestra: <|emotion:enthusiasm|>And there's a transfer result that interests me. One magnetic sensor model hadn't appeared in the representation's pretraining, and the authors report using the frozen encoder on it without sensor-specific encoder training.
Eris: What still needed training or adjustment in that case?
Vestra: The downstream controllers still underwent task training, and the sensor input needed the right geometry and normalization. <|prosody:expressive_high|>That's evidence of transfer to that tested sensor, not proof that an arbitrary new hand or object will work without adaptation. These are the authors' results, not an independent replication.
Wrap-up
Eris: You can find the sources and follow these stories at Ground Truth, <|prosody:pause|>groundtruth dot day.
Vestra: <|prosody:expressive_high|>What would you want to know about how an AI tool handles your data? Tell us in the comments, and follow AI in Detail for more news and research. <|emotion:contentment|>Thanks for listening.