we just released our first model: SID-1 it's designed to be extremely good at only one task: retrieval. it has 1.8x better recall than embedding search alone (even with reranking) and beats "agentic" retrieval implemented using all frontier LLMs, including the really large and

We improve both pass@1 AND pass@n during training. The issue is that lots of claimants: 1) train on domains with heavy mid/posttraining in the base models (math) 2) don't train for very long In many of these small-scale experiments, gains come from re-learning the format

Label noise really matters in RL SID-1's task requires reporting the documents most likely to contain the answer to a question. When the ground truth data contains errors, the model will start overreporting in hopes of catching spurious targets. For one public dataset where

Most RL frameworks are fundamentally unstable. We wasted more H100 hours on debugging this than any other issue fornour multi-turn, multi-env RL run (below). When using OpenAI-style messages for env interactions, parsing and retokenizing leads to subtly different tokens. This

Good RL environments are much richer than you think. We evaluate training for 100 epochs and see eval reward increase steadily. Partly, this is because our RL setting allows obfuscating the answer between epochs, largely mitigating memorization (when inspecting train rollouts).

We believe retrieval is the ideal playground for self-play RL on LLMs. SID-1 was trained with "pseudo self-play:" New questions were generated to cover gaps in model behavior as training progressed. We think we're not far away from closing that loop: Generating hard, verifiable

SID-1: Tech Report It has way more detail than is prudent. https://t.co/GOdo8HFBh3

mediocrity is contagious and terminal

computer use and code gen progress is outpacing general intelligence improvements. why? you can easily create synthetic data for both. let me explain: if you have *more* high-quality data on a task, a model trained on that data will be better at it. currently, that data is

three letter AI companies that start with S are really having their day

give me an nvidia-smi -L long enough and i shall move the world

AMDAHL'S ARGUMENT FOR AI The productivity speedup AI apps can provide is limited by how much human-in-the-loop work is required. Humans are ~1-3 tokens per second. They can't really be sped up – unless you're @neuralink. So if your application requires a human completion for

we call this a "Context Engine" internally. you want a service to give the model definitions and extra information about the request at hand, without forcing users into providing it manually every time. humans are really good at inferring context from past interactions, the

"Uncertainty routing" is the real news in the Gemini announcement. Without it, GPT-4 still beats Gemini in CoT@32 on MMLU! For people building apps: GPT-4 is still better in zero-shot. (charts from the Deepmind Gemini Technical Report)

.@karpathy love the pricing for the vision API! great way to reduce our @openai bill at @try_sid

I still firmly believe the race to functional AGI is about adding personalized context about you, not better models. Hopefully that’s a focus for devday.

.@daltonc with the shades. yc reunion with @sama was fun yesterday.

FYI you can add personal apps & services to any LLM (Llama 2, @OpenAI etc.) with @try_sid using a single API. If you want to see OpenAI + apps & services (through SID), you can try the ChatGPT plugin we stealth launched last week (already has +500 users). Unlimited Drive &

Exactly what we're building @try_sid! But not just for one bot - but for EVERY AI tool out there! 😉

Here are 8 key considerations for building *production-grade* LLM apps over your data (RAG) 💡 (see 🧵): 1️⃣ Chunks used for retrieval shouldn’t necessarily be the same as chunks used for LLM synthesis (@md_rumpf) 2️⃣ Embeddings should live in a different latent space than what