Synthetic data is the accelerator of the next phase of AI — what it is and what it means.

When what seems like pure LLM black magic is actually supported by the literature.

Databricks’ new model is surpassing the performance of Mixtral and Llama 2 70B while still being in a size category that's reasonably accessible.

Mamba, StripedHyena, Based, research overload, and the exciting future of many LLM architectures all at once.

Huge steps forward in confirming that RLHF can really help you on vibes based evaluation, among many other RLHF analyses.

A PSA everyone needs. The importance of a wait and see attitude when it comes to new models, big and small, open and closed.

Looking at the difference between two sets -- what rumors say industry leaders are doing with RLHF and what the literature is up to. I'm starting my new series studying RLHF literature.

Listen now | The first Interconnects research interview! We go even further on the promise of state-space models in the emerging LLM market.

Llama 3 shows that scaling won't be a limit for open LLM progress in the near future.

Most of the arguments about "safe" releases of open LLM weights are nearly dead in the water.

What the details tell us about the most in-vogue LLM evaluation tool — and the rest of the field.

Modeling the compute versus performance tradeoff of many open LLMs.

Now we will have some grounding for when weird ChatGPT behaviors are intended or side-effects — shrinking the Overton window of RLHF bugs.

A sampling of recent happenings in the multimodal space. Be sure to expect more this year.

Basic tips on how to assess inbound ML content and cultivate your news feed.

Models that seem totally out of scope from recent open LLMs give us a sneak peek of where the industry will be in 6 to 18 months.

A proper critique of the Foundation Model Transparency Index (FMTI). Plus some thoughts on the ecosystem implications.

Where moats are tested now that so many people have trained GPT4 class models. Claude 3, Gemini 1.5, Inflection 2.5, and Mistral Large are here to party.

Direct (DPO) vs. RL methods for preferences, more RLHF models, and hard truths in open RLHF work. We have more questions than answers.

Evaluation is not only getting harder with modern LLMs, it’s getting harder because it means something different.

We have an amazing open mixture of experts model for the holidays!

Integrating some non-computing science into reinforcement learning from human feedback (RLHF) can give us the models we want.

The state of the ML communities big and small starting 2024. My general expectations for the year.

There are plenty of jobs, but finding a place where you're happy is as hard as ever.

At ICML, it is obvious that many people are getting value out of RLHF. What is limiting the scientific understanding of it (other than research embargoes)?

A small model at the beginning of big changes.

Emergency blog! Three things you need to know from the ML world that arrived yesterday.

Where does releasing model weights fit into company strategy? 3 requirements, 3 actions, and 3 benefits of being in the open LLM space.

Certain definitions of AGI are backing people into a pseudo-religious corner.

Emergency special: The information we need to understand what Q* is was right in front of us, but the memes are more fun than reality.

How does autonomy of language models relate to data collection.

Modern LLMs are becoming the easiest and most efficient way to access information. This will change how we see the world.

A proposal for a new definition of an “open-source” LLM and why no definition will ever just work.

In an era dominated by direct preference optimization and LLM-as-a-judge, why do we still need a model to output only a scalar reward?

The coming image-generation battle and its implications on ML product longevity.

Multimodality may be a key differentiator in how moats are built for LLMs.

The community reaction to Llama 2 and all of the things that I didn't get to in the first issue.

The cutting edge technical discussions beneath the wow factor.

Google rejoins the open model party and gets some backlash for a frequent problem for generative AI.

Scale’s making over $750 million per year selling data for RLHF, who’s coming to take it?

And why the comparisons don't really matter. Repeated patterns in the race for reproducing ChatGPT, another year of evaluation crises, and people who will take awesome news too far.

Arguments for and against rapid progress in robotic learning research.

Failure modes on the quest to general, open-source LLMs. Expect pivots to specialized models.

Many, many signs of life for preference fine-tuning beyond spoofing chat evaluation tools.

The deployment path that’ll break through in 2024. Plus, checking in on strategies across Big Tech and AI leaders.

An interview I've wanted to bring you for a while.

This fact is why so many debates around LLMs feel broken, especially moderation.

When is GPT4 going to schedule my meetings? What is stopping it?

The core themes of ML and the blog this year. What changes in 2024.

How to read what multiple people mean by the word openness and see through the PR speak.

A fun demo on how generative AI can transform content creation, and tools for my fellow writers on Substack!

We need to heal AI research norms, not build a super project. The reflections and reverberations that we're feeling in the AI community from Oppenheimer's quest for the atomic bomb.

Meta is continuing to deliver high-quality research artifacts and not backing down from pressure against open source.

Culture wars, open letters, new politics, developer days, and everything hidden under the smiling face of RLHF.

A list for beginners and wannabe experts and everyone in between.

Two stories will begin to unfold as the AI capabilities-to-product overhang is reduced.

Definitions from open-source software are being bent by new machine learning technologies.

Why changing the dynamics around the distribution of malicious content is a growing concern in addition to just the generations themselves.

SF continues to be the center of attention for developments in AI, but this time it's in the physical world.

(some buggy audio in this one, from MoE rather than Mixtral lol)Mixtral: The best open model, MoE trade-offs, release lessons, Mistral raises $400mil, Google's loss, vibes vs marketingWe have an am...

No, SemiAnalysis, HuggingFace isn't misleading all of open-source, and open-source is still making real progress.