I've defaulted to Ubuntu/Debian for so long for Linux stuff that I'm out of my wheelhouse on other distros, but I needed a newer kernel and was having trouble with mainline so I switched to Manjaro. Manjaro has been absolutely great so far! Try new things out y'all!

Nuremberg-style trials if we make it through this as a country

I really dislike "unalive" - why are we fully capitulating to some old bad classifier? Say the word dammit! It's not even a swear word, or I guess it is now?

It took too many data points (their published articles) for me to understand the Thinking Machines trajectory, but it's very cool what they seem to be building up to. I really hope tinker becomes an industry standard. Sanity! And let billions of adapters bloom. thinkingmachines.ai/blog/announc...

With all these new different architectures coming out - Qwen3-Next and DeepSeek V3.2 and Meta CWM and all - I feel bad for the llama.cpp maintainers. I wonder if a new local inference engine will spring up soon. Maybe it's Modular/Mojo's time to shine.

In gradeschool we each got a 5-digit number to type when we got our lunch. A few years later I started mixing that pin code into passwords. Years later still, I switched everything to a password manager - but decided to put that number into the unlock phrase. So I remember my 3rd grade lunch code 😂

People complain about TikTok or Instagram brain rot, but in my experience it's Snapchat that's the worst offender. I get the absolute worst tabloid spam push notifications from that app. Literally tabloid celebrity stuff. What a shame.

Turns out that if you throw 5-15 million tokens at a question, you can get a really good answer

Don't bother with the Meta Ray Ban Displays. I drove down to Meta, got to my appointment, and they told me I have to come back, in person, AGAIN, to purchase. What was the point of this apointment then? Waste of an afternoon. Just wait until you can buy them online.

Too many interesting ML papers right now! This is great though.

Update: it's going away 😢

If OpenAI wants everyone to switch over to their new responses API, they should contribute implementations to vLLM, llama.cpp, and SGLang. Otherwise chat completions will always be the lingua franca. developers.openai.com/blog/respons...

It is probably something that can only be experienced, but as someone exiting his 30s, it is fun and fascinating to see people take on their true second acts. Open source code collaborators become artists, entrepreneurs, influencers. Interns become powerful CEOs. People surprise and impress you.

Holy shit I just realized that "chat" as gen-z/a uses it is innovative in that it solves the gender neutral plural issue. "Hey guys" / "Hey girls" / "Hey chat". Cool!

We are so far past the rubicon

Rewatching S1 of Heroes and it's still a good season of TV. It will be fun some day when AI can re-imagine a show e.g. branch from S1 and take it in a different direction than the real S2+

This post should go around again for a new generation. Using some of these new ML systems and I'm reminiscing, thinking about how much hard-won distributed systems knowledge is probably lost to the erstwhile walls of Eddie Rickenbacker's @jmhodges.bsky.social www.somethingsimilar.com/2013/01/14/n...

Hyperparam sweepin' - first time trying AdEMAMix (runs 0.2.2, 0.2.3, 0.2.4, 0.2.5) - it's good!

Foundation finale: wow

Started this model training yesterday BTW. Custom 500Mtoken post-training dataset that I distilled from GLM-4.5 Air (single turn), K2-Instruct 0905 (multi-turn and instruction following), and gpt-oss-120b (tool calling) - should be a drop-in replacement for Qwen3-4B-Instruct-2507

Qwen3 Next looks super interesting! It's actually incredible to me the way Alibaba is perfectly executing the Llama strategy Zuck laid out last year. They are really out-executing. I am rooting for team 🇺🇲 so I hope we find a way to catch up and surpass! qwen.ai/blog?id=e34c...

Just so sad about my country, and the world right now. Love one another.

It must be so hard to keep an Anthropic-sized GPU cluster going with high availability and usable for research etc. That said, they are making big mistakes at a time when competitors are fiercer than ever and devs impatient status.anthropic.com/incidents/k6... status.anthropic.com/incidents/72...

I remember having an idea earlier today that I thought was so good. I then remember thinking to myself, "I should write that down so that I remember it later!" I then remember thinking, "nah, this idea is so good, how could I forget?"

For some reason I just thought of this viral "dystopic internet" image from 16 years ago, went and found it, and it's so interesting. I thought it hyperbolic at the time. Now it seems quaint! Sentiment from back then: www.reddit.com/r/pics/comme...

Three co-equal branches of government, my ass.

Trying to wrap my head around this one. You don't hear much from ASML but IYKYK they are central to semiconductor manufacturing. EU-strategic investment, maybe? www.reuters.com/world/europe...

Worth pointing out that "Community Notes" have seemingly failed on all platforms. Mis- and disinformation is more rampant than ever and lies do seem to spread faster than truth. Latest one to get me (that I know) was that wonderful Marlon Wayans poison apple comment - unfortunately completely fake.

I think part of Anthropic secret sauce might be that they probably work hard to make Claude a fantastic (superhuman really) prompt engineer. Which has all kinds of downstream effects, because Claude prompts itself all the time, especially in agentic workflows. And seems easy to throw RL compute at.

The Cold Start Is Over: Social Products in the Age of Agentic LLMs gist.github.com/ericflo/d2ba... - this is the second time I rambled into a voice note (this time 11 mins) and used Gemini to transcribe and GPT-5 to turn into a blog-ish post - it's nice to get these ideas out! Poor md render though.

I have a visceral negative reaction to the certification part of this. Can't fully articulate why yet, but seems somehow dangerous (from a worker power dynamics perspective) to have this all from one entity.

app.fireworks.ai/models/firew... Really, really hard to beat that price vs. token quality

Our info environment is wild right now, harder and harder to determine epistemic truth

Is there some sort of viral marketing hack going around telling people to position their products as expensive, or something? Every other Instagram ad I see starts like "People always complain about our prices, but our prices are high because it's quality..."

What about this idea? An AI treasure hunt. First you put together a website where entering the correct password gives a $200 gift certificate or some prize. Then you hide the password in the training data of an LLM. You also hide several decoy passwords.

Earlier today I rambled a 12 minute voice note, had Gemini transcribe it, then GPT-5 rewrite it into a blog post-ish format. Out pops "Start With the Judge: A Practical Blueprint for Dataset Engineering" - not bad! gist.github.com/ericflo/f49e...

I always thought some DeSantis or Cotton type politician would find a way to capture MAGA and sideline Trump, but over time I've come around to the idea that Trump is unique with his cult of personality. I now believe that without him, the right reactionary movement recedes for a while and infights.

It honestly really sucks to be the frog boiling in the water now, when I've been pointing out the rising heat for a decade. "Sideloading" was my big issue last decade due to being blocked by Oculus back in 2016, and we are now in a far worse position. Tilting at windmills I guess. But won't stop.

If Google goes forward with this, I will be a hater for at least a decade and probably for life. We need to find a way to punish them if they proceed. It simply *cannot* be allowed. 9to5google.com/2025/08/25/a...

The coding agent is Just Like Me (v2_best_really_final)

Something kinda fun about this Rich Sutton talk, he seems fired up by that conference, said he was up 'til 4am working on the talk 🤣. Love the push towards simplicity - that the big idea should fit on a slide. youtu.be/gEbbGyNkR2U?...

Just realized everyone who's been using Claude Code has a local directory full of frontier-level agentic rollouts in ~/.claude/projects/*/*.jsonl tailored specifically to their own codebases and preferences. [1/2]

LLM RL seems to trade off ICL for IF - a bummer IMO

With the advent of vibe coding, I want to see a TV show that's like Project Runway but making apps/sites/games instead of clothes. Celeb guest judges, 24-48 hours per challenge, the whole thing. Unconventional challenge is something like, you have to target Internet Explorer 5.01 on Windows NT 3.51

brookechan.notion.site/OpenAI-API-C... I have this bookmarked from the GPT-3 days. Fun to go back and see how many of these early explorations have become full product categories!

Remember Trac? trac.edgewall.org

Oh no. I think I had a good idea.

Someone did it! xcancel.com/jxmnop/statu...

It's hard to know what to post during the literal, actual, downfall of America.

OK gpt-5 generated this pretty good satire about itself! 🤣