ml , local llm enjoyer
🇨🇦
Joined July 2022
- Tweets8.9K
- Following1.1K
- Followers2.1K
- Likes14.7K
Pinned Tweet
BrainBox v2: muscle memory for Claude Code.
It learns your repo from the agent's own habits, then predicts the next file and tool before the agent goes searching. Runs locally in SQLite, no LLM calls.
New paper, tested on 137 real agent sessions. Full thread:
this bug is so old it is more likely a backdoor purposely build by agencies
no way a bug like this survived since 2010
fable 5 ? without distillation and from a non Chinese lab 🎉🎉🎉
We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.
It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2 weeks.
for those having issues applying rea on binaries via claude
aka refusals from opus then start the project with a chinese model and once you face any issue or have made like 40
percent progress then involve opus
git isn't the archaic part. a content-addressed DAG scales fine.
what breaks at 30k merges/day is everything bolted on top. PR review, CI queues, merge locks, all built around a human clicking approve.
agents don't need a new VCS. they need a new merge layer.
In February I built BrainBox, a memory for Claude Code that learns your repo from its own habits.
Today: v2, with a proper eval on 137 real agent sessions.
It predicts the agent's next tool better than every baseline. Here's how:
5/ Files are harder. BrainBox ties simple recency there, and on my own projects a plain frequency count scored higher.
Why: only 17% of the files an agent needed had been touched in any earlier task. Most files are new to it.
Tools are habits. Files are tasks.
6/ Next: a live on/off study to see whether better prediction actually makes real tasks shorter.
Paper: zenodo.org/records/23250505
Code: github.com/thebasedcapital/b…
npm i brainbox-hebbian
lmao already announced post claude’s , this new alliance between @spacexai and @AnthropicAI is so much more thrilling
cal ai went from zero to $50M arr in 18 months with no vc money. one person running the pipeline replaced an entire growth team. variant volume beat creative genius.
THE CO-FOUNDER WHO RAN GROWTH AT CAL AI JUST LEAKED THE ENTIRE AI UGC PIPELINE. ONE PERSON, ZERO CREATORS, THOUSANDS OF TEST VIDEOS
Cal AI went to 15M downloads and a $40M year before MyFitnessPal bought it.
He co-founded it and ran their UGC and influencer programs the hard way, and now says this system would've made him 10x more productive.
• The loop
> Claude Code plus Apify pulls the outlier videos in your niche, the ones doing 12x an account's normal numbers.
> Claude builds a local gallery, you pick what to clone, an analyzer breaks down hook, tension and payoff by timestamp.
> Higgsfield generates the AI presenter from a real creator frame, Seedance 2.5 animates it, one test take before any batch.
• The part that makes it an army
> Finished exports go straight into Postiz over MCP, one draft per creative per platform.
> Claude writes the captions, proposes the schedule, you approve, it publishes to TikTok, Reels, Shorts and Facebook.
> Each platform reports back, winners get remade with real creators, losers die cheap.
The whole thing runs in one Claude Code window -> research, generation and distribution with no human in the loop until approval.
He's handing out 13 prompts and the full skill file. Article below ↓
bertrand meyer says ai can't write software because it just generates text that happens to compile.
the field's actual history is copying stack overflow answers and tweaking them until they work.
the ai skipped the browser tab.
sicpers.info/reasons-to-disl…
