Search: “llm”

Reasoning from scratch, round number 5! This time, talking about log-probability scoring (also a great fundamental concept for loss functions like cross-entropy in pre-training and distillation) and s

A blank cell can change the meaning of a forecast. Four times a year, the Fed's 18 top policymakers each put their forecasts on paper: where growth, jobs, inflation, and interest rates are headed. Thi

Automated AI theorem proving has moved the frontier: The holy grail is not an AI that can produce an endless pile of true theorems. It’s an AI that can discover mathematics, and build each discovery i

btw i had no idea what i was doing when i started this but learned so much about models in the process and feel like i have a much better sense of what’s happening under the hood managing someone (an

Pay attention to this new wave of System One models if you are building custom harnesses. First Jev. Now, Contrastive Language Model (CLM). CLM is 9x faster than Jev. CLM seems to be a better verifier

Today we establish that AI is as good as expert human tutors for immediate GRE learning gains (p=.015, n=2,383 students, study conducted July-Sep 2026). One hour with an expert GRE tutor: $75. One equ

There’s an enormous amount of open training data on Hugging Face. What does it take to make it work together 🤗? For Marin’s 535B run, we built on 25T tokens from 152 datasets with licenses permitting

Claude Opus 5.5 is the first Opus variant that can automatically fall back from its main model to a reduced‑capability version when executing a small set of specialized tasks associated with frontier

Great paper from Google and colleagues. Trains Text-to-SQL agents using multi-agent RL. (bookmark it) This work proposes DualSQL, which splits Text-to-SQL into two agents, one that links the question

Impressive paper showing how much the harness changes a coding agent's results. Harnesses do play a huge role in what you are getting out of the models. GPT-5.5 was run inside both Claude Code and Cod

Interesting work on long-horizon research agents. PrimeScientists can decide where a research agent spends its budget. Achieves 10.3% more reward with 50.6% fewer research attempts, under the same bud

First model of Laya was produced with a single overnight training and subsequent fine‑tuning run, taking only 15 hours from the start of training to publishing the model. It was released on Hugging Fa

Document classification and splitting idea using @typesafeai Jev based on LiteParse, and LlamaParse. Classify: categorize probabilities, and review flags. Measure: separate OCR and decision timing, pr

Found a great production use case for Jev. I used Jev to organize ~2.3K AI research papers. The total cost was $0.14, and it took about 83 seconds. The process: The papers already had old tags, which

StepFun’s new Step 5 Preview model is impressive! Had a chance to test it early. I've been testing it as a coding agent. It’s on the Pareto frontier for cost vs. capability. It is a very capable model

Thanks for the model, we were able to port Laya to coreml with 99.5% of the ops on ANE + benchmarked too. it is now blazing fast with 3.7 ms per decision on an M5 Pro. Release: https://github.com/Flui

the reason i'm excited about decision models: agentic systems are just good engineering around models. 'build prod, not god' these types of fast, cheap & accurate models are super important for that t

Introducing DocJev - a lightning-fast OSS library for document classification and splitting with jev ⚡️ Give a document alongside some natural language category rules. Jev will predict the document ca

Nos travaux déplaisent à certaines petites frappes. Celui-là est candidat à l’élection présidentielle 😅 Cédric O, ancien ministre du numérique d'Emmanuel Macron, détient 90 millions d'euros en portefe

Banger paper from MIT and Sakana AI. They show that self-improving coding agents work. The best part is that their approach, Self-Improvement via Fast Tree-search (SIFT), runs at a tenth of the CPU ho

Cédric Villani après l'annonce de la solution d'OpenAI au problème du millénaire : « J’ai été secoué. Une ambiance de fin de l’histoire. C’est un cataclysme comme jamais les maths n’en ont connu. » Cé

Inference scaling part 1. Starting with a modded text generation function (temperature scaling, top-p filtering, multinomial sampling) to generate diverse outputs for self-consistency and best-of-N (i

I've seen people describe Jev as an 'AI if statement'. But what if it actually WAS an if statement? Introducing Probably: a programming language powered by Jev: https://probably-lang.southpolesteve.wo

Super interesting work from Zoom and colleagues. If you maintain a hand-built coding harness, there are some great insights here. (bookmark it) They held the execution loop of a coding harness fixed a

Epoch AI has launched “Benchmark Reviews,” an initiative to evaluate existing AI benchmark datasets. The program initially audits 15 benchmarks, categorizing them as 4 verified, 9 flawed, and 2 lackin

Perhaps the most important warning I ever issued was this, to the US Senate, May 2023: “A month after GPT-4 was released, OpenAI released ChatGPT plug-ins, which quickly led others to develop somethin

Banger paper from NVIDIA. It's on the topic of choosing which models go into a multi-agent system. The team compared eight selection strategies, based on size, accuracy, answer diversity and error div

Nice work from Adobe. It's standard practice to store a fixed reference answer for every eval case. When the underlying data changes daily, that stored answer goes stale. Adobe researchers write each

Agents are coming for video production. @PooldayAI is going for this. You prompt it once, and it edits your real footage, picks the right generative model for each task, and checks its own work before

full support this proposal from @DKThomp, and there should be a similar broadening for negligent defamation Drop Site and parts of the antitrust left are poo-poohing the idea that we need new rules to

🆕 Humanity’s Last Invention — with @RichardSocher! https://www.youtube.com/watch?v=eDFXtSg3zB8 @recursive_si is the latest neolab to burst on the scene with a $5B fundraise and one of the most impress

Introducing our first agent tool: the Brand API This helps gen AI agents solve 3 problems: - Extract: produce on-brand assets when users already have a brand - Search: find and retrieve brands for ins

Some food for thought when designing benchmarks... So, here's a little computer-use (visual) comparison between GPT-5.6 Astra and Qwen3.8 Max. The task here was to recreate the image in the center usi

Well.. @levie and I filmed this episode of Training Data a week or two ago, when the “current thing” was Doug Leone’s novacaine root canals instead of pacing the frontier… Simpler times! But Aaron’s a

You are what you eat and AI models especially so https://x.com/EMostaque/status/2099580512675262512?s=20 The Labs need to focus on 10x or 100x their effort on ALIGNMENT. Stop training the models on th

I completely agree with this take related to the “AI will create viruses that will kill us all” BS. But let me add my few cents, because I am really really angry! I worked with one of the deadliest vi

Live now: our Harness Engineering Track from AI Engineer World's Fair 2026. Ports, proofs, kill switches, and what is left of an agent when you remove the model. Thesis: when an agent breaks in produc

Excited to continue our partnership with the @alibaba_cloud team on this one, bringing the latest version of Qwen's flaship model to the Fireworks community. Start building with Qwen3p8 today: https:/

Interesting paper from Microsoft. If you run persistent memory for a production agent, this one is worth your time. (bookmark it) Memory curators usually read only the finished trajectory. That lets t

TLDR (written by AI because I can't be bothered to write short pieces for short attention spans. Original piece written with my own two thumbs): Today’s frontier models can’t “copy themselves like a v

Very cool paper on memory compression for agents. If you run many agent sandboxes in parallel for RL or evals, memory becomes highly redundant. This work suggests that compressing against that redunda

Another brilliant paper from Meta. This one is worth reading if you work on production-grade ranking or recommendation systems. Meta deployed an autonomous agent that runs the ML iteration cycle acros

If you ask a coding agent to write a program that maps x to y given some (x, y) examples, you're doing a very expensive form of symbolic learning. In fact what you're doing is specifically a form of D

Machine learning doesn't end when the model is trained. You still need to evaluate it, deploy it, and build the infrastructure to keep it working in production. Today's Deal of the Day takes you acros

a lawyerly, evasive, short-sighted response. we didn’t “directly” use the data but yeah we might’ve trained on it who the hell will ever want to collaborate with them after this? Two things to disting

Banger paper from Google DeepMind. Every AI weather model so far has been trained and initialized on analysis data, which is itself the output of another model. That means the forecast inherits whatev