Periodic Labs reduces inference costs with Kimi

When Jev is confident, it can agree with more powerful models at a fairly high rate. In one my of tests, it agreed 95% of the time with GPT Luna 5.6, at about 200 milliseconds a call. It classified around 12K Claude Code interactions into 16 kinds of work (bug fix, new feature, planning…) in 11 minutes; Luna took over an hour. Here are a few notes on that and other brief experiments. https://lnkd.in/gZB7MPhu

New Video Published! Atomic state management in React with Jotai, addressing Context + Hooks bottlenecks Link to the video in the comments

Extra thinking need not mean longer chain-of-thought text. Looped flows show tiny recurrent models can “think longer” by integrating a probability flow — Sudoku climbs from 74.5% to 97.9% as steps rise — without writing more tokens. Full Note: https://lnkd.in/eqNt6zvW If this was useful: more Notes like this (and the daily Five things digest) free at https://lnkd.in/ekGFSzDN — subscribe if you want them in your inbox.

We're back to algorithm work baby, let's go! First step is allowing the live reshape mechanism to pull discoveries the moment they ingest. userTasteResponsivenessLevel ++

In our benchmark, Jev matched Claude Opus 5 at 87% hallucination-detection accuracy while running 23x faster and at roughly 1/300 the cost. But our first run told a very different story: Jev scored 76% vs. Opus 5 at 83%. The difference came down to one line of code: the decision threshold. That's just one learning from our benchmark where we ran 23,325 judgments across RAGTruth and SummEval, comparing Jev, Claude Opus 5, and GPT-5.6 Terra on accuracy, cost, latency, and calibration. Get the full findings: https://lnkd.in/g-CNKkJM

Generative Pre-trained Transformers (GPTs) are great at, well, generating. But this superabundance of content can very quickly become problematic. The problems of too-muchness, such as piling-up backlogs to read, mushrooming reports to approve, and unaccounted-for repositories of unvetted content, all put a massive strain on our cognitive system that leads to overload. In this environment, the premium skill is actually discernment: knowing how to delay, decay and destroy this generative content. Yes, suddenly the skill we all need is de-generation. One good example of how this could be done comes from the world of insects. Stigmergy is a phenomenon that ants and termites use to manage their collective societies in a way that keeps them going. The environment itself is the medium of communication, telling all the participants where to build, where to maintain and repair, and where to decay and destroy. With AI in play, I believe organisations have something to learn here. But it starts with us: only individual "de-generative" discernment will open us up to this kind of organisational architecture.

Breeze TTS 2 beats ElevenLabs on Artificial Analysis and can design voices from text prompts. We tested it locally and found one catch: the license makes it hard to actually ship.

Shiva.AI 2.0 is ready. We’ve taken Shiva.AI a step further and turned it into a framework, making it easier to integrate with your own code and build on top of it. For a simple test, we put Shiva through one of the most famous control problems: the inverted pendulum. Using just 5 nodes, Shiva controls it successfully. Shiva.AI 2.0 is now ready to use, test, and verify and get back to us if any issues persists. We’d love for you to try it and tell us what you build. Link for the repo: https://lnkd.in/gEtkH23D

What happens when you let agents play on a playground? This was a single experiment with 44 trials. I learned, agents tend to be deterministic from a starting point unless you give them memory or force a random seed. Agents, left to their own playground, use their tools: create and edit text and symbolic strings and write code. When you have a multi-agent system, a lot of computation time is spent with agents reading each others' logs. I also did a simulation of a single agent on a playground and it was boring. What experiments would you do next? Hitendra Chaturvedi Thomas Kull

Close to 1000 ML engineers have downloaded our guide to DINO, DINOv2, and DINOv3 since we published it. 🔍 That number means a lot to us, because DINO is not just a topic we wrote about. We were one of the first teams to actually integrate it into a product, so the guide comes from real work rather than a summary of the papers. If you want to go deep on self-supervised learning and how DINO, DINOv2, and DINOv3 fit together, the full guide is here: https://lnkd.in/eFXpGNR4