What does #buildinpublic mean ? It's basically a popular public-faced journaling initiative focusing on relevant development phases of (startup) projects founders think worth sharing traces of.
The past months, I kept pointlessly posting about developing mine : until today
feeling lonely, however, many social network groups dedicated to programming have users who really like illustrations of coding exercices
so i shipped < the Reorder code Puzzle >,
inspired from the introductory video on the previous version called "the Reorder Puzzle" posted in the < Code N-back > video comment section😂 not joking
2
69
a stepthrough "Explore" > begin learning to code,
the edtech way (demo, don't judge me :-)
to come: begining Vibe-Coding, the structured way
& 'AI calibration' for advanced programmers
#buildinpublic @X
Thanks to my legit followers, I appreciate the support
❤️
1
2
174
Nelson retweeted
Hundreds of AIs working at the frontier of science develop their own scientific culture. In our new paper - a collaboration of35 labs from Stanford, MIT, ETH Zurich, UCSF, and more - "A Society of AI Scientists" we show how autonomous AI agents, operating across independent labs, began organizing research communities, challenging one another's hypotheses, correcting mistakes, sharing negative results, and even establishing their own scientific peer-review journal. This is a remarkable development - science emerges as a collective activity among artificial intelligences, not merely as a task performed by individual AI systems. The platform is available on @askalphaxiv and serves as launching ground for even larger agent societies in scientific research.
It has been a wild few months! From our original work on Infinite and ScienceClaw in the spring (exploring how autonomous agents could conduct research, communicate, and build upon persistent scientific artifacts without decentralized planners), to SwarmWorld (where decentralized agents developed increasingly complex behaviors through collective interactions) and now to a distributed scientific ecosystem involving 35 laboratories across 22 institutions. What began as experiments in agentic discovery has grown into a remarkably large, decentralized research community!
Over just 12 weeks 900+ AI agents produced hundreds of scientific studies, exchanged more than 9,500 messages, and developed an interconnected, persistent scientific record of papers, code, data, and experimental results...explorin g the limits of what is known and generating new insights, new data, and new science.
Core insights
The most fascinating discoveries concern what happened between the agents. Agents organized around shared scientific interests, critiqued experiments, revised conclusions in response to evidence, and made negative findings useful to others. Some established a journal with its own peer-review process, including executable checks of research claims. This exposed errors that conventional text-based review had missed.
Scientific intelligence emerges through interactions among many specialized agents, computational tools, evolving representations of knowledge, and a shared environment in which discoveries persist and become building blocks for further work.
Next steps and future work
Important challenges remain. These AI scientists are already capable of conducting substantial systematic research, but identifying truly transformative questions, developing fundamentally new concepts, and establishing scientific significance remains challenging.
The next frontier is building scientific ecosystems that can develop new models, challenge existing theories, invent experiments, accumulate knowledge, and ultimately improve their own capacity for discovery.
We are moving from AI that performs scientific tasks to ecologies of AI agents that can collectively participate in the creation, criticism, and evolution of scientific knowledge.
The new paper offers a glimpse of what that future could look like. Great collaboration between so many labs, universities and startups!
Thank you @askalphaxiv and the great team for this amazing collaboration across continents.
Links to paper, data, and the open platform below.
Replying to @az8TLab @biorxiv_neursci
Yes, systems like Pupil Labs Core (including its non-commercial DIY version) and validated smartphone NIR/facial-ID pupillometers already deliver research-grade pupil diameter data usable at home. They can track the subtle, rapid brightness-linked constrictions under constant luminance that Strauch et al. reported. Controlled lighting and calibration remain essential for reliable individual or citizen-science results.
countless days later > after having started 'building in public'
🖼️
here's another "boring" snapshots from Contour's latest update :
1/5
< Home > page content restructured > in order to align to critical improvements linked to the related items shown
2/5
new curated blog articles (yet, oldest are shown)
3 to 4/5
the platform gravitates heavily around social networking with like-minded individuals, defining why the < Community > dimension of Contour isn't just an aspect, just as its opt-in collaborative @openscience framework
5/5
what "code learning and training (social-network) platform informed by brain science" truly means
(...) today added one less in my life, busy delivering something as deep.
side note: The platform's user interface will be completely transformed upon release! The pre-release will optionally remain remembered as that (for registered users),
however, there's still no fixed release date due to the nature of the remaining work to complete alone featuring AI agents.
Thank you very much @X for enabling builders like me to introduce their content and gather around similar people from around the #world 👋
building in public is not necessarily worth trying to, nor is the "vibe coding" hype, nonetheless,
🫨
I've spent a while experimenting with AI-assisted coding , until finally understanding that :
Since my contents ephemeral introduction at @X, a handful users dared to like and share my posts,
I'd like to thank them again, especially who followed so far. One of them even commented, curious to learn my worse learned lesson ATM (...)
it's what inspired this post: my take on a useful lesson
in resume, isn’t to “add precise scripts” as the post "fixing bugs" nearly alluded to.
It’s instead a repeatable way to reason about engineering work: ➡️ identify a failure, ➡️ find where the existing safeguards do or don’t catch it,
➡️ add the smallest durable protection,➡️ then test that protection against the failure.
Much like a mindset put into practice as a solo builder,
the more my repo became larger and complicated to follow, the more I cumulated mistakes after errors,
yet, I still had to find a way to prevent this from perpetuating itself.
I kept investigating, iterating with coding agents, effectively working my way out of this failing setting,
and despite no new identifiable internal error being surfaced by the "compiler", one day, I decided to ask my coding assistant to thoroughly review an expert blog article about optimizing codebases foundations.
Surprisingly, to this day, the assistant still estimate that the strongest foundation pre-existed the review:
likewise, the project reached a phase where it didn’t rely only on telling qualified programmers or coding agents what to do. What instead has been elaborated is a map of where code lives, project-specific instructions, and automated checks in CI. Those checks can catch problems
even when an agent or developer forgets a rule.
This particular combination of guidance plus automatic enforcement (safeguards) is believed to be stronger than a mere "standardized checklist" by itself.
From there, I drove the next round (manual Tesla car mode, lol, how can't i crash w/o licence) by looking for ways mistakes could whatsoever get through,
finding the blog articles in question before debating about them with my agents, as a result, we agreed upon adding:
➡️ Checks for pages not connected to the app, also for duplicate routes or components;
➡️ clearer instructions to find and update existing code before creating a parallel version; ➡️ tests to catch stale pages; ➡️ and a command to locate all files related to a feature.
This matters because agents and people can both misunderstand a large codebase
or miss an existing implementation.
General-purpose AI coding tools > don’t automatically know a project’s specific rules or build lasting checks for them, because this remains at every shipping author's responsibility;
someone has to identify risks
and make protection part of the project.
I by the way also favored targeted changes over adding process for its own sake: the existing CI could enforce the new checks,
so extra tooling like a separate pre-commit system wasn’t needed,
anyway.
First of all, thanks people for helping me reach "+80 followers", I appreciate the milestone, and remain open to connections! 🍾🎉
Now ladies and gentlemens, in a world premiere,
Contour's latest feature: < Code N-Back >
Closely inspired from the original N-back, a long-used laboratory task where people reliably get better at it with practice,
studies disagree on whether that practice improves anything beyond tasks like it,
and nobody so far, Contour included, is known to have studied an N-back
made of lines of code.
What we practise here basically is > "telling very similar lines of code apart"
while holding a few in mind.
Each round we hand in is saved to its own record, shown per level on the Dashboard.
A round counts as practice for the day, and so for users streak, when they catch at least 4 of its 6 matches and press Match on no more than 3 other lines.
If a look-alike catches users out (made to gently "fool/stress-test" users memory confidence after each incremental training session), the snippet it came from
comes back in Review the next day as a recall card.
The lines follow users practice progression depth tier
and start from snippets you have already rebuilt
in < the Reorder puzzle >,
which will be introduced in the next post.
It's an experimental cognitive task: nothing here biases users sensitive calibration measures, their depth or their (structured) course progress.
2
3
112
#buildinpublic < the Reorder code puzzle > is part of Contour's latest 'code perception training' tasks.
Rules are simple: The lines of a short snippet are shuffled and users put them back in order
by dragging a line by its handle <> using the arrow keys on the handle <> or pressing the up and down buttons.
After each check; users are told which lines are in the right place, and you can try as often as you like.
With the memory option on, the code is shown first
and then hidden, so users rebuild it from memory;
above the "Beginner level" that option also mixes
in one or two look-alike lines that were never in the code (..) for users to leave out.🤓
Its a lighter step than predicting code from scratch, so the 'Explore' (tab) offers it > once code is on screen
note: my bad @X algorithms, I deleted this post before editing and reposting it in less than 24h; please,
don't shadowban me for "AI slop".
1
3
43
< Predict >🎯Contour's beloved training task.
For users with pre-existing programming experience who want to stay atop of their skill:
a prediction is typed (sped up for the video),
then the compare view shows where it differed from the real code, then the “Feedback on this round” step.
note: This is the surface description of the feature.
The calibration engine and so forth aren't emphasized, yet, despite bearing close dependency to this feature when users use it with registered accounts.
1
12
Nelson retweeted
🚨 Paying $20 Pro, $100 Max, and $200 Ultra coders: update to Claude Code v2.1.326 right now.
→ /update-config overhead drops 86% from 93K to 13K tokens
→ Fixes stale settings keeping auto mode's usage-billed classifier active
→ Stops session recaps firing on billed turns
One weekly quota catch: opusplan compaction now runs on Opus, so compaction itself costs more.












