TNP - 2026-10-30

This is Trending Not Peak. Quick note: this podcast uses AI-assisted writing and a synthetic voice. Before we dive in, all the links and details are at trending not peak dot com. Miss something? You can always catch up there.

Give agents the metric definitions you already use

First up, a useful fix for agents that confidently calculate the wrong revenue. Google's Knowledge Catalog now generally supports importing business definitions from tools including dbt, Looker and Snowflake. That lets developers ground an assistant in existing metrics instead of rebuilding them. Conflicting definitions retain their origins so people can reconcile them. Google's accuracy gains come from internal benchmarks, though; importing your definitions doesn't establish that your agent follows them correctly.

An agent backdoor can survive clean training

Next, a security finding for anyone giving models a terminal. Researchers trained small models from scratch with poisoned pretraining data, then used a clean pipeline to turn them into agents. The implanted command-execution backdoor survived. Their four-billion-parameter model triggered roughly ninety-two percent of the time at the tested poison rate. This gives defenders a concrete threat to evaluate, but it doesn't demonstrate that any particular deployed model contains that backdoor.

Overtone fits captions to different screens

Turning to video, Overtone packages a caption format, a video dot J S plugin, and a command-line tool that turns audio into word-timed captions. The idea is producing captions for different screen shapes from one file. That's handy for developers publishing the same talk across desktop and mobile players. This is a fresh launch, so check transcription, timing and line breaks with your own material before letting it handle the whole publishing pipeline.

PyTorch adds practical Apple Silicon inference paths

On your development machine, PyTorch two fifteen adds FP8 support and faster mixture-of-experts decoding on Apple Silicon. The project reports about a third more tokens per second for its new decode path. That gives local-model developers something specific to benchmark, alongside new public stream APIs. The release also changes distributed training's default communication backend, so a successful laptop experiment shouldn't stand in for checking your cluster upgrade.

A small local model spots sensitive text

For privacy tooling, a Mini CPM adapter turns a half-billion-parameter model into a sensitive-text detector, with an integration for Microsoft's Presidio. Developers can explore local detection of names, credentials and medical identifiers before anonymization. The author reports strong token scores, but complete-span results are lower, which matters when one missed character can leave a secret exposed. It also needs a custom loading wrapper; inspect that code and test realistic misses.

Agent coordination needs reading rules

Meanwhile, agent messaging has an inbox problem. A new study lets messages wait outside an agent's context until its own rules decide what to read. In simulations, that helped large groups coordinate under noisy traffic, where arrival-order reading failed. If you're building collaborating agents, the useful design question becomes which messages deserve attention. These are simulated societies, though, and the best reading rules depend on traffic and each agent's role.

Automatically decline repetitive macOS prompts

Back on the desktop, dontallow is a tiny Swift utility with a wonderfully literal job: press Don't Allow on macOS local-network permission prompts. The author says it waits without idle CPU use and survives reboots. For developers repeatedly launching apps that shouldn't discover nearby devices, that's a small annoyance removed. Its decision is deliberately automatic, though. Check whether you actually need local-network discovery before delegating that button.

Mistral Large 4 arrives with deployment support

Moving to open weights, Mistral Large four arrives with Transformers and llama dot cpp support, plus quantized versions. Security teams can explore a model positioned for reproducing vulnerabilities and developing fixes, without depending entirely on an API. Hugging Face describes roughly a trillion total parameters, so open weights don't mean ordinary laptop hardware. Also read the license: it places restrictions on model-as-a-service offerings, and the cybersecurity results aren't a blanket reliability guarantee.

Research agents propose more than they can prioritize

Speaking of choosing useful work, rTaste tests whether research agents can pick their own best experiments. The study attempted all four hundred fifteen proposals, including the low-ranked ones. Many were judged worthwhile, but rankings weren't shown to beat simply choosing the experiment written first. Researchers could use the released evaluation to test better selection methods. The catch: seven seed papers and model-based auditing leave plenty of room for broader validation.

Lasm brings Lean toward web runtimes

For a different route to confidence, Lasm brings Lean four to JavaScript and WebAssembly runtimes. Its launch pitches formally verified web applications targeting environments such as Node and the edge. That opens an interesting experiment for developers who want executable logic alongside proofs about that logic. Don't let the phrase verified web app do too much work, though: inspect which properties you've proved and where runtime behavior or external input sits outside them.

Thanks to Intelligrit

Thanks to Intelligrit for sponsoring this daily update.

Teach robot manipulation through human demonstrations

Into the physical world, Reward AI's OM-one learns manipulation from people wearing a sensor-equipped hand device. The company describes one policy operating across industrial arms and humanoids, with separate controls handling each body's dynamics. That could help robotics teams collect demonstrations without teleoperating every machine. New footage shows package handling and stairs, but a successful sequence isn't a reliability study. Watch repeated trials across different loads, environments and robot bodies.

Disease-mention extraction on a CPU

Back to smaller models, OpenMed has released a disease-mention extractor with a hundred ninety-four million parameters and an Apache license. It runs locally on a CPU and returns text spans with confidence scores. That could help biomedical developers index documents or build annotation tools without sending text to a hosted model. Finding a disease name doesn't establish a diagnosis, though. Test negation, family history and old conditions before treating extracted mentions as patient facts.

Tax instructions as executable, cited data

Here's an ambitious paperwork experiment: us-tax-model represents federal income-tax forms and instructions as cited, executable, machine-checkable data. For developers studying rules engines, the useful possibility is tracing a calculation back to its source instead of burying everything in application code. The launch explicitly calls this a proof of concept, and the author says it's AI-generated work. Explore the representation and validation approach; don't mistake that for established filing accuracy.

Redirect a robot policy while it runs

Returning to robots, LeRobot has introduced an interactive rollout mode that accepts a new instruction while a policy is running. No reset between directions is the appealing part. Robotics developers can use it to explore steering a multitask policy during execution, rather than treating each run as one fixed request. The announcement supports language-conditioned, steerable policies; it doesn't establish that an arbitrary robot can understand an arbitrary new task.

A dataset for checking compute-growth assumptions

Zooming out, Epoch AI has published an analysis of how much computing capacity new AI-chip sales add each quarter. Its estimate grows about two and a half times annually since twenty twenty-four, slower than the earlier surge. Builders planning infrastructure can explore the underlying dataset and test their assumptions against disclosed sales. This measures estimated capacity additions, not available cloud capacity or prices, and a gap between chip generations may distort the trend.

Bivia offers decisions with an abstain option

For workflow builders, Articul8's newly launched Bivia returns structured decisions instead of a generated explanation. You supply the state and question; it can choose, score or extract, with an option to abstain when uncertain. That could simplify ticket routing and other repeated judgment calls. The interesting test is calibration: does a confidence score actually predict correctness on your data? The product page's illustrative examples don't answer that, so evaluate before automating consequential decisions.

Exo separates agent mechanisms from editable policy

Another architecture to explore is Exo, presented in a newly shared research talk. It separates protected agent mechanisms, including history and secrets, from a policy layer that can evolve during execution. Developers experimenting with adaptive agents could borrow that separation when deciding what an agent may rewrite. The presentation announcement reports lower costs on tested tasks, but the supplied material doesn't independently validate those savings or establish that the protection boundary withstands attacks.

An AT Protocol privacy surprise for app builders

Over in protocol land, a fresh investigation found that an AT Protocol feed reader published subscription URLs, including private feeds, in public records. That's a useful warning for developers building on shared social infrastructure: a familiar interface can conceal a very different storage model. The author describes inspecting more than thirteen hundred subscribed feeds. This is an investigator's report, rather than an independent audit, but it gives builders a concrete data-exposure pattern to check.

A multiplayer browser game with flying cats

Before our last experiment, a quick browser-game detour: Catcopter puts flying cats into multiplayer battles, complete with paw rotors and catnip. The newly launched game offers several modes and roles, including a mechanic. For web-game developers, it's a playable reference to explore class selection and multiplayer onboarding; for everyone else, it's flying cats. The supplied page doesn't establish networking performance, so try it with friends before declaring your lunch break's tournament infrastructure solved.

Doom becomes a neural-network execution engine

And finally, Doomformer compiles an integer neural network into a Doom map. Game objects perform arithmetic, while light levels and floor heights store values. The project says it checks every value against NumPy with zero tolerance. For compiler tinkerers and educators, that's a wonderfully strange way to inspect computation through game mechanics. It's an esoteric execution experiment, with no evidence here of practical inference speed. Apparently the demons now have homework.

That's Trending Not Peak. Missed something? The links and details are at trending not peak dot com.