I love Google AI overviews and the haters are wrong (even if they’re bad for Google’s business).

Thinking, Searching, and Acting A reflection on reasoning models. It's easy to fixate on the "thinking" that gave reasoning models their name, but just over a year out from o1-preview's release by OpenAI, the core primitives that make up models today has expanded. buff.ly/BpUShgg

Claude Code works better in poor internet than OpenAI Codex CLI. The former can work on a plane with meh wifi (maybe not perfectly), the latter couldn't even get started.

Whether or not these drive meaningful value and are seen as substantial steps in AI abilities or robustness seems crucial for the continued ramp of AI capex.

Core reasoning model pieces rn 1. Reasoning tokens -- thinking 2. Search -- acquiring relevant information 3. Code -- building Most other tools are just a form of code or search. All of these relate to in-context learning.

What is the best AI voice assistant for learning Chinese, is it ChatGPT (where I'd try it first). Has anyone tried this? Any advice?

My 2nd pikes peak ascent today (13.3mi, 7800ft of gain), 3hr 19min (7 min pr!). Great event. Just one hill for a donut. Would recommend.

Would be an incredible own goal. We should be making it easier to bring top talent from abroad not harder.

Rumors going around are that NeurIPS PCs had to reject ~400 papers because of venue constraints that were in the accept pile. Hard things like this happen but the PCs and conference should explicitly tell all the authors who had this happen to them. Or create something like accept w/o presentation.

Explaining how to get the most of the CLI agents and why they're so important to understand today and tomorrow's AI progress. They're showing the new fundamentals of agents and how frontier labs will hill climb on open-ended tasks. buff.ly/jAAhHnQ

I finally got around to making a tool to compare completions from SFT vs. RLHF trained models. This is a mini site for the RLHF book that I've wanted for a while. buff.ly/lqDL5wa It's always been hard to say what RLHF does to a model within a more complex post-training pipeline.

SemiAnalysis had this awesome plot in the latest post, but now we need to know where Google stands:

Cool new codex behavior -- it uses a preview software to take screenshots of the front end it's writing, to debug visually instead of just code. Then deletes the screenshots and makes changes or wrapping up.

Word on the street was that Codex CLI was currently beating out claude code, and this was before the model update -- GPT-5-Codex today. Excited to try it out! Still figuring out my first use case for Codex the agent, but the CLI is much better now. buff.ly/yCcIRBt

~9 months ago I spent some time with Luca Soldaini making a list of models and resources for language models that were more than just open weights (data, code, logs, etc included). It's getting out of date, could use some community contributions :)

Getting ready to invest more time into the RLHF book to prepare for print edition. What do you wish was clearer or had more coverage in it?

Great podcast overviewing recent robotics work, foundation models, trade-offs, ideologies, limitations, etc. buff.ly/AwAo31W

Crazy random line in the Cursor RL blog post saying they're collecting RL data from real users, updating the checkpoint every 90 to 120 minutes. This would've been unthinkable if you told me it 12 months ago. What a wild year of reasoning models. buff.ly/8LqreS2

Just got invited to attend/speak at an event where the names aside me highlighting the event were affiliations of AI Envoy to the Pope and The Onion. If this isn't peak bubble it's peak something.

Latest open artifacts (#14): NVIDIA's rise, "Swiss & UAE DeepSeek," and a resurgence of open data While Qwen takes some rest, others continue to fuel the open model space. buff.ly/9MKUgYR 43 of the best models/datasets from 27 different organizations.

Latest sign of AI company affluence is the Gemini team sending premium Fellow water bottles and Peak Design satchels to influencers like me. The AI bubble is well and strong. Guess I’m ready for the next Gemini release now

Developers & researchers can literally "pay to win" in their career right now by paying for better AI models. Those can afford it are trading a bit of money now for a faster career progression rate and more compensation later. $200/month is cheap relative to most people's long term earning potential

Threads is my platform with the most numerical followers (almost 60K now) but it feels like a total ghost town. No substance — bizarre really.

I'm using GPT 5 Pro a lot these days, mostly for research and information processing (not much for coding yet). It's the most consistent, smartest model I've found and that's totally worth the money.

If the AI lawsuits get any bigger the tech companies will just buy the publishers instead of paying them (same goes for Reddit et al)

The weirdest VC subsidizing of our time, 10% of the Anthropic series F goes to writers

Beautiful afternoon on Mt Sunapee

ACL 2026, a top tier NLP conference, being over the July 4th weekend is so bizarre (4th 2026 is a Saturday). I'll be turning down any more talk invites in lieu of touching grass. People should probably consider submitting elsewhere.

COLM is coming up! Very excited. I'm starting to figure out two things: 1. A small invite-only dinner for Interconnects AI (Ai2 event news later). 2. Various research chats and catchups. Fill out the form below or email me if you're interested :) 🍁🇨🇦 Interest form: buff.ly/9nWBxZ9

Feels like the first time in weeks my follower count here has gone up 🤣. Lack of growth and churn is still blue sky’s biggest issue.

My vacation is winding down but ai is still on a bit of a late summer siesta

Pretty big vibe shift on open models coming from a predominantly AI Safety oriented org to say "gatekeeping access to general-purpose technology is not a sustainable or proportionate response to low-confidence evidence of serious risk." Pretty much my point for a few years.

Got the essentials, not much more is needed.

Mt. Lincoln & Eisenhower

On vacation for a bit, this is now a dog and mountain account until usual programming resumes. Ps I really wanna put Phoebe on the cover of the RLHF book.

The AI Action Plan marked a policy pivot toward open source as a way to accelerate AI. What's that mean in practice? @natolambert.bsky.social (@ai2.bsky.social) & Keegan McBride join @alanrozenshtein.com & @kevintfrazier.bsky.social to provide their answer. podcasts.apple.com/in/podcast/t...

Just signed a book deal for The RLHF Book, excited to make improvements to it this fall and get physical copies in your hands soon :) (rlhfbook dot com)

I made a snarky Western list to show how far behind we are in fostering an ecosystem vis a vis China.

China's Top 19 Open Model Labs We ranked all the organizations in China releasing open models, from the top of DeepSeek to small, newer academic labs making waves with tech reports and niche models. buff.ly/YLSBBAq

Claude Code w Opus and GPT-5 Pro are worth the money.

GPT 5 Pro feels like what my hope for o3 pro was (that it didn’t deliver): robust, detailed, search model. Using it whenever I want something like deep research but for an answer rather than a report.

My top 5 most memorable models from using them at/soonafter launch: 1. Claude 3.5 Sonnet (personality, all round perf) 2. o3 (search behavior + perf) 3. o1 pro (robustness) 4. Gemini 2.5 pro (long context + perf) 5. GPT 4.5 (personality)

I disagree with Dwarkesh Patel on continual learning being a major bottleneck for the current path of AI. It may be a bottleneck on making a "more efficient and human like AI" but language models are well on track to adapt quickly and precisely to personal work.

The Open Source AI Question | @natolambert.bsky.social of Ai2 and @robertwrighter.bsky.social cover the waterfront on open source AI—including its pros and cons and the implications for US-China dynamics and, btw, the meaning(s) of ‘open source AI’.

I just stumbled across the fact that I've published 300 posts of some form on Interconnects (or under its previous names). We'll see if I make it to 1000. Would be fun.

Zuck confirmed Llama 4.1 and 4.2 coming soon in the last earnings call, despite all the superintelligence rumors of stopping open models. I've personally heard rumors of a Llama 4 8B, which makes sense given the success of 3.1 8B. (screenshot from Stratechery)

I wanted to provide more color on this because its a bigger deal than the main Ai2 account is hyping it as. For example, the entire annual budget of the NSF for AI in 2026 is $655M dollars. To commit to a training models on single line item of about 20% of that is a huge deal.

Open model performance rankings across most benchmarks these days. Top half: China Bottom half: Everyone else So many such cases.

If we want to build good open models we need to have well-funded labs dedicated to it. It's easy to get excited about community projects/ decentralized training but those aren't ready to make good models. Those are great for research ONTOP of good models.

Model router is prolly the new news feed. Everyone says they hate it, in a few years people love it. Model picker is cursed idfk about that.