Dmitry Noranovich retweeted
Claude writes better code when it can verify its work. Checks that used to be expensive, like tests, benchmarks, and end-to-end runs, are now cheap.
Here’s how the Claude Code team thinks about verification:
claude.dev/blog/how-the-clau…
Dmitry Noranovich retweeted
Hey everyone! I am super excited to share our new research report! 🎉
Scaling Reranked Retrieval 🚀
Thread with more details! 🧵 (1/17)
Dmitry Noranovich retweeted
What is the role of Reranking in Agentic Search? 👾
Search Agents usually use the same retriever on each turn. ♻️
"Question's Gambit" argues that the first search deserves more: an AI search system with query decomposition and reranking. 🪄
The original question is decomposed into clues and each clue gets its own BM25 search. These results are pooled together and fed to a reranker. The top 5 from this stage becomes the agent's first observation.
Then the agent runs its normal loop with plain BM25, no reranking after the first turn. 📝
On BrowseComp-Plus with GPT-5.5, accuracy goes from 83.1% to 90.5% with about the same number of agent-issued searches. 📈
What do Rerankers and Chess have in common? ♟️
I am SUPER EXCITED to publish the 150th episode of the Weaviate Podcast featuring Ghita Houir Alami (@ghita__ha), Co-Founder of ZeroEntropy and now leading Model Research at Notion! 🔥
This is one of my favorite deep dive podcasts into Search! 🔎
We explored how pairwise preferences can be turned into useful pointwise scores with ZeroEntropy's zELO method! Like chess Elo, a sparse sample of pairwise battles between documents is enough to fit an absolute score through a Thurstone model, and that signal is then distilled into a pointwise reranker. The result... 🥁 ... the state-of-the-art zerank-2 cross encoder! 🔥
The podcast covers a lot of interesting topics in Search! I loved learning about Reranking with Jev / Decision Models. 🛠️
Ghita explained how many teams want cross encoders to prune and classify rather than just order results. Calibrated relevance scores can support searching with a Dynamic K, thresholding, and classification. 🧠
We also discussed the tradeoffs between pointwise, pairwise, and listwise ranking. ZeroEntropy's pairwise reranker was slightly more accurate, but distilling it into a pointwise model came close to matching it at much lower inference cost. ⚡️
We also covered infrastructure for rerankers, agentic search and reranking, instruction-following rerankers, metadata in training rerankers, and the relationship between RSI and SLMs! 📜
This was a super fun conversation, and I hope you find it useful! 💚
YouTube: youtube.com/watch?v=mQgYPtct…
Spotify: spotifycreators-web.app.link…
Dmitry Noranovich retweeted
Introducing Gemini 4 Hydrogen, an extremely capable and fast model built to bring Gemini 4 to everyone!
Available to all tiers of Google AI Studio, Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro + Ultra subscribers in Google Antigravity.
Today we’re introducing UI SDK.
Agents are becoming both the builders and the users of interfaces. So we asked ourselves: where should the abstraction live?
Take a Select. There’s a popup, selection logic, keyboard navigation and focus management. Those same behaviors are useful in lots of other components.
So we pulled those pieces apart, tested them, and made them reusable. The behavior stays the same, whichever framework, renderer you build with or the harness generating the UI.
That’s the idea behind UI SDK. Give agents the pieces to generate the interface you need, right when you need it. Even interfaces we haven’t thought of yet.
Dmitry Noranovich retweeted
introducing Gemini 4 Hydrogen
this new model brings cutting-edge intelligence to everyone — with speed and efficiency for everyday coding and agentic workflows
start building with Hydrogen via the Gemini API and in AI Studio: aistudio.google.com/prompts/…
Dmitry Noranovich retweeted
Make predictions on tabular data without training a new model for each dataset.
In Tabular Foundation Models for Data Science Agents, built with @prior_labs, you'll give an AI agent a tabular foundation model as a prediction tool.
Enroll for free: bit.ly/4sMtNOz
Dmitry Noranovich retweeted
Announcing /swarm, a new slash command in Google Antigravity for orchestrating specialized agents in parallel across the desktop app and CLI.
Turn a single prompt into a coordinated multi-agent workflow where agents plan, build, and verify together on a live message board, with the flexibility to mix and match different models for each role.
Dmitry Noranovich retweeted
Local AI 301: Quantization www.teapawt.org/i/broadcasts/2rJXOgajp…
Dmitry Noranovich retweeted
Gemini Robotics 2.5 is here. 🤖
Our latest physical AI is able to learn hundreds of everyday skills - from folding laundry and serving pastries to autonomously sorting car parts on the factory floor.
Here’s a snapshot of its capabilities 🧵
During pre-release testing, Claude Fable 5.2 and Claude Mythos 5.2 worked on ambitious projects in many different fields. Here are a few.
From a one-paragraph prompt, Fable 5.2 designed three working electronic devices, including this home voice assistant hub.
Dmitry Noranovich retweeted
What do Rerankers and Chess have in common? ♟️
I am SUPER EXCITED to publish the 150th episode of the Weaviate Podcast featuring Ghita Houir Alami (@ghita__ha), Co-Founder of ZeroEntropy and now leading Model Research at Notion! 🔥
This is one of my favorite deep dive podcasts into Search! 🔎
We explored how pairwise preferences can be turned into useful pointwise scores with ZeroEntropy's zELO method! Like chess Elo, a sparse sample of pairwise battles between documents is enough to fit an absolute score through a Thurstone model, and that signal is then distilled into a pointwise reranker. The result... 🥁 ... the state-of-the-art zerank-2 cross encoder! 🔥
The podcast covers a lot of interesting topics in Search! I loved learning about Reranking with Jev / Decision Models. 🛠️
Ghita explained how many teams want cross encoders to prune and classify rather than just order results. Calibrated relevance scores can support searching with a Dynamic K, thresholding, and classification. 🧠
We also discussed the tradeoffs between pointwise, pairwise, and listwise ranking. ZeroEntropy's pairwise reranker was slightly more accurate, but distilling it into a pointwise model came close to matching it at much lower inference cost. ⚡️
We also covered infrastructure for rerankers, agentic search and reranking, instruction-following rerankers, metadata in training rerankers, and the relationship between RSI and SLMs! 📜
This was a super fun conversation, and I hope you find it useful! 💚
YouTube: youtube.com/watch?v=mQgYPtct…
Spotify: spotifycreators-web.app.link…
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
Dmitry Noranovich retweeted
Catching up on the 10 (!) open LLM architecture releases from the past month. Quite a lot happened in October.
I guess we can call it "Open October"!
1. Kolibri-1 (78B): A notable model from Germany. Features classic Grouped Query Attention (GQA) and Sliding Window Attention (SWA) attention module.
2. Ling 3.1 Flash (560B): Uses Kimi Delta Attention and the eighth uses gated Multi-head Latent Attention (MLA) inspired by DeepSeek.
3. Muse Spark Open (816B): Besides Musse Glimmer, the first major open weight release from Meta (since Llama). Uses classic GQA + SWA but adds a LatentMoE.
4. Nemotron 3.5 Super VL (120B): Combines Mamba-2 layers with just eight attention layers. Also LatentMoE
5. Dory1.0 (4B): A looped transformer design.
6. Reflection Beam (501B): GQA+SWA architecture.
7. Recursive Flint (40B): Note that "Recursive" is the company name; the architecture is a fairly standard MoE with GQA and SWA.
8. MiniMax M3.1-Flash (428B): Uses their own MiniMax Sparse Attention module.
9. LFM3 (1.7B): Hybrid architecture that mixes short convolutions and attention.
10. Mistral Large 4 (1T) combines MLA with sparse attention.
Will probably cover/discuss things in more depth in the coming weeks!
Dmitry Noranovich retweeted
Scaling Discovery requires Forgetting! 🌊
I am SUPER EXCITED to publish the 149th episode of the Weaviate Podcast featuring Shangyin Tan (@ShangyinT), a Ph.D. student @UCBerkeley! 🔥
Long-running optimizers stall when an agent remembers too much. Forgetting can get them exploring again! 🏔️
The story starts with GEPA and its successor, `optimize_anything`, which extends these ideas from prompt optimization to code, GPU kernels, and agent harnesses! 🚀
Shangyin explains why prompts and harnesses are really a generalization test, whereas a single kernel is pure optimization. I loved this nugget for understanding how optimization algorithms are extending from models to code. 💡
Running GEPA, Autoresearch, and Meta-Harness as backends behind `optimize_anything` resulted in every one of them plateauing a couple of hours in.
As sessions grew, agents went from editing 100s of lines to 5 or 6 due to being stuck in a local optima. Handing only the best candidate to a fresh optimizer and forgetting the history got progress climbing again. 📈
We also discussed how to design Multi-Agent teams, the use of Git as a communication protocol for Multi-Agent AI, and how to handle expensive evals! 💸
This was a super conversation, I learned so much from Shangyin, and I really hope you find it interesting!
YouTube: youtu.be/Y21NgZvx3Rs
Spotify: spotifycreators-web.app.link…
Dmitry Noranovich retweeted
We've raised the max request size for Messages and Token Counting on the Claude API from 32 MB to 128 MB.
Image-heavy requests like long computer-use sessions and big PDFs were previously hitting 32 MB with plenty of context window left. Now you get ~4x more room.
Dmitry Noranovich retweeted
Introducing Gemini 4 Carbon, it's a big jump over Argon which we released just 2 weeks ago.
Gemini 4 Carbon is SOTA on cyber, long horizon tasks, ranks #1 on Artificial Analysis V5, significantly improves on usability in domains like coding, and is more aligned.
























