Everything I publish, in the order I published it: dispatches from Twitter, videos from YouTube and notes from my vault, archived here in full — text, images, video and links — so they outlive the platforms.
while google is busy renaming notebooklm to gemini notebook, a chinese open weight startup completely outclassed every gemini model on the market today by a long shot.
how can this possibly happen?
@signulll Different understanding of scarcity, maybe?
@FonziAlerts@hiiinternet@ycombinator@Accel@a16z Thank you for sharing, I didn't have Fonzi on my radar. Same for us, a small company. Hiring engineers is absolutely still top-priority, there's budget etc. Even in a post-AI world, finding good talent is still the hardest part I find.
I let a genetic algorithm loose on the geometry of a 7-inch quadcopter frame — the real, open-source Source One V6 plate drawings, morphed by fourteen genes and flown through six simulated weather scenarios, with Claude sitting in every few generations to propose designs from the run’s own telemetry. The bottom line: we evolved the champion to fly 16% more efficiently (Wh/km) across six weather scenarios.
new post on harness engineering for AI self-improvement: t.co/ZYvGfVs61k
It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple.
Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.
@lilianweng Yup! I believe auto-research loops will be part of it. And kind of are already. I experimented with a Karpathy-style DIY research loop on a personal quest and, yeah, this stuff works like today.
And heeeere (?) is a reply that's really just a line of text in a markdown file. (My goal is to also post replies from Kern.app because, tbh, going on Twitter to give meaningful replies is, to me at least, a losing battle to too many distractions.)
I wanted a beautiful app to explore ocean wildlife. Fable built this in an hour. It generated videos with Seedance and carefully synchronized them to make these absolutely insane transitions. I've never seen anything like this in an app. Unreal. t.co/W1NjAnlzQu
Wow. I have to say this is actually uniquely beautiful and engaging (and I'm usually the first one to nit-pick on one-shots with a "yes but..." )
I'm curious what's going to happen next: You're the first one I've seen with this type of result. I'm genuinely amazed.
How will I feel about the second and third app? And the hundredth in 5 months? Will this new level of picture-perfect quality eventually become part of my AI-filter?
Alright! "Loop Engineering" is the term we're using? Works for me. Here's my latest loop engineering project. Built on @NimaSiboni's recent "The Heuristic Scientist: Open-Ended Algorithm Discovery with LLMs" workshop at @TUBerlin.
I think my take on it came out nice AND I found some nifty new levers that proved to be super useful, especially the auto-pivoting facility.
"Auto-Pivoting on patience"
Give the agent a patience — how many rounds it may go without a new best. Spend it, and it must pivot: drop the current line for a genuinely different groundrule. Our main guard against rabbit-holes.
"Runs that build on runs"
A search isn't one fixed run. Each new run inherits the champion — the best groundrule so far — and climbs from there, while you change the model, the patience, the library around it.
"Hand-curated inspo"
The library is an optional, human-written input — notes, framings, even a personality handed to the agent before it thinks. It doesn't fix the answer; it shapes the frame the search starts from.
"Run it on your own hardware"
The loop talks to any OpenAI-compatible endpoint, so the proposing model can run locally — LM Studio, or anything serving the same API. Point it at your own hardware and the search runs unmetered: no rate limits, no quota walls.
Yesterday I organised Europe’s first community Cafe Compute with @cerebras in Berlin!
We had over 200 signups and ~80 people join us to talk about fast inference and running frontier intelligence at the speed of 1000+ tokens per second.
Special thank you for Cerebras and @communidiyi for entrusting me with the Ambassadorship and @WigetJanette with @aicampusberlin for hosting us and helping me organise this evening.
In the first half we had Zigfrid (Solutions Architect at Cerebras) join us and answer all the questions we had about what they do and how their chips work.
And in the second half we had great demos from the community:
@slobkebap presented a real time GenUI flow to build UI in realtime by talking to it using @OpenAIDevs's realtime v2 api combined with Cerebras GLM-4.7
@SpringStreetNYC presented his autonomous AI research lab experiment in which he tried to teach a small model to ask great questions on the level of frontier models.
Hoang Minh showcased how he finds vulnerabilities in big OSS projects very fast and helps patch them.
And my personal favourite demo that night: @leolurch, who showed us his machine learning pipeline to find lost art pieces on auction platforms. Basically face recognition but for art, to recover potentially stolen or simply lost art from the last centuries.
I had so many cool chats and am glad that so many very talented people showed up.
Thank you everyone for coming!
And thank you so much @0xSero for sharing the event!
Me, trying to touch the future while giving a quick demo of "'Which city in Paris are you staying in?'–Can an LLM autoresearch loop teach a phone-sized small language model (SLM) to ask frontier-quality questions? Can it? Does it? Let's find out!"
Hey @LumaHQ! You app is absolutely fantastic. In only three weeks, I've been to five top-notch events meeting new, totally cool people.
But now that you've solved meeting new people for (thank you), I have a NEW problem:
As the events start to wrap up, we often want to stay in touch. Cool! Yay!
Now, to "connect-connect", the norm is to use LinkedIn. So, now I've to go back to (shudder) LinkedIn as well. You know, that place. And it's super-duper awkward: their app has a QR-code contact scan thing that gets more embarrassing every passing second while you're both fiddling with your phone ("...oh, doesn't work, there's a glare on your screen. Oh, it's my camera?...")
After yesterday's event (awesome btw.), I seriously consider printing contact cards. But then, what do I put on there? My phone? My email? My Twitter? And then?
What actually HAS worked a few times: I meet them, memorize their first name, then go to Luma, check the participant list (when it's visible) for their first name, when there are multiple "Davids", deduce based on first impressions which one and then try to find them online. Somehow.
So, team Luma, as a huge fan: I would like to ask you to let me find and connect and stay in touch with these new peers via your app. Thank you so much :)
Right now and somewhere, the next Kane Parsons/Jony Ive/Steve Jobs is graduating from/dropping out of college. Given today's demos, they're CLEARLY NOT anywhere near SV but I'm sure you can find them.
PLEASE do and put them on those same $840k/year salaries. Give them a year or so. Let them define the next two decades IN and ON their own terms.
Airloom“The snuggle is real” — evolving quadcopter frame geometry for Wh/kmA genetic algorithm with Claude as an occasional co-designer bred a 7-inch quad frame 16% more efficient across six simulated weather scenarios. Every candidate explorable in 3D, with real OpenFOAM streamlines. Free software, GPLv3.↗ alexisrondeau.me
alexisrondeau.me"Which city in Paris are you staying in?" - A post-training study on small-model specialization with an autoresearch loopAcross 77 experiments — every lever from prompting and retrieval to fine-tuning and 2025's rubric-grounded RL — an on-device 3B model landed within about one rubric point of a frontier cloud model on every quality dimension: close, but consistently a notch below. None of the post-training playbook closed the last notch, and the controlled tests show why — the gap is a broad capability limit, not a…↗ alexisrondeau.me
GitHubGitHub - akaalias/kern: The text editor I've always wanted: The beefy love-child of Emacs, iA Writer, and ObsidianThe text editor I've always wanted: The beefy love-child of Emacs, iA Writer, and Obsidian - akaalias/kern↗ github.com
UNKNOWN PLEASURES RECORDSShe's Lost Control, by JOY DIVISIONfrom the album Bains Douches Paris 1979 remastered complete version (free CD version out of print)↗ hivmusic1.bandcamp.com
