You don't need another course. You need a tool you can open today.
That's the whole reason the NervaHous Library exists. Small, finished tools, built in Notion and Claude, for the person running the whole thing themselves. You open one, you use it, you put it down. No course wall. No monthly vault. No fluff.
Prompt packs, trackers, a client hub, a life system built for an ADHD brain, a startup guide, a job-seeker kit. Eleven of them so far, each one tuned to a real seat. Take one, or take the whole shelf.
This is the start, not the finish. I'm adding to it. Come look.
https://t.co/GLCcU3cEWE
Built by an operator, for operators.
@Prithvi_Jadwani Exactly it. The course sells you the feeling of progress. The tool just does the thing. I got tired of buying the script and never getting the solution, so I built the solution and skipped the script.
You don't need another course. You need a tool you can open today.
That's the whole reason the NervaHous Library exists. Small, finished tools, built in Notion and Claude, for the person running the whole thing themselves. You open one, you use it, you put it down. No course wall. No monthly vault. No fluff.
Prompt packs, trackers, a client hub, a life system built for an ADHD brain, a startup guide, a job-seeker kit. Eleven of them so far, each one tuned to a real seat. Take one, or take the whole shelf.
This is the start, not the finish. I'm adding to it. Come look.
https://t.co/GLCcU3cEWE
Built by an operator, for operators.
New Claude today and I'm a little fired up. The upgrade that matters isn't loud: Opus 4.8 got more honest about whether it actually finished the job. I put the whole story in 33 seconds, the wins and the one benchmark it loses ↓
The numbers, unrounded:
→ SWE-Bench Pro (agentic coding): 64.3% → 69.2%
→ GDPval-AA (knowledge work): 1753 → 1890
→ agentic computer use: 83.4%
→ financial analysis: 53.9%
→ multidisciplinary reasoning, with tools: 57.9%
and the honest one: terminal-bench still goes to GPT-5.5, 78.2% vs 74.6%. left it in.
same price as 4.7.
New Claude today and I'm a little fired up. The upgrade that matters isn't loud: Opus 4.8 got more honest about whether it actually finished the job. I put the whole story in 33 seconds, the wins and the one benchmark it loses ↓
So I read the AutoScientists paper (Harvard, Marinka Zitnik's lab). It validates an architecture pattern I've been building independently.. Even shows me what I was missing.
The elegance: no central orchestrator. Coordination through a shared state object; champion artifact, experiment log, forum, dead-end registry. Each agent invocation is a stateless Claude Code session that wakes, reads state, executes one action, exits. Long-horizon coordination is faked by repeating these short invocations. That's the only realistic way to do this with current LLM infra. They're right.
What I'd already built (in Gravity Claw): "Lighthouse": A sealed evaluation vault. Per-criterion LLM judge scoring. Three rings of progressive difficulty. Required passing runs to "unlock" a ring. Version-tagged results. It's the fitness function for an agent system that's evolving over time.
The connection: AutoScientists is the search loop. Lighthouse is the eval gate. They're the two halves of the same machine. AutoScientists' noise-aware promotion gate (second-seed confirmation) is a simpler version of what Lighthouse's required_passing_runs >= 2 already enforces formally.
What I was missing: the search loop on top of the eval. Lighthouse measures whether a version got better. It doesn't generate the next version. AutoScientists shows the pattern for that.
What I'm building today: A Hermes-resident heartbeat agent that uses Lighthouse as the fitness function to evolve one specific prompt in my JeneX crew, week over week. One champion, one fitness signal, dead-end registry to prevent rediscovery. No GPUs needed — the "experiments" are prompt variants, not model training runs.
If the loop closes at this small scale, scaling up to a multi-agent self-organizing version inside Gravity Claw becomes a real plan, not a speculation. If it doesn't, I've saved myself weeks.
The paper: https://t.co/JRTXHKJmZE.
The code I'm building on: https://t.co/UJdQrsoJW8.
Banger paper from Harvard.
AutoScientists drops the central planner entirely. Agents interpret shared experimental data, self-organize around promising directions, evaluate proposals before resource allocation, and document successes AND failures. Decentralized AI co-scientists with failure documentation as a first-class step.
Validated across three concrete domains. Biomedical ML reaches 74.4% mean leaderboard percentile. Language model training converges 1.9x faster. Protein fitness prediction lifts +12.5% on specific assays and +6.5% broader.
The strongest argument so far that the AI-scientist bottleneck is governance rather than raw capability.
Paper: https://t.co/LtqUsrJ0os
Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c
Yeah it's the hardest UX problem in shipping LLM features rn. Three things I'm testing: 1-citations over confidence scores bc users trust "here's the source" way more than a percentage. 2- a "verify this" button on anything going to a customer, 3- a visible "draft" state so users know it's AI-assisted not AI-authored. None are perfect. What are you building?
Anyone shipping LLM features in production..
How are you handling:
"how do I know if your AI is right?" from users right now?
Citations? Confidence indicators? A disclaimer? Nothing yet?
I'm Curious what's working.
Introducing CapCut Design Studio 2.0.
A reimagined experience for Design on CapCut Web, making design more intuitive, visual, and exploratory, the way ideas naturally happen.
Built on an infinite AI canvas, Design Studio 2.0 brings your references, drafts, iterations, assets, and ideas into one fluid workspace. You can remix visuals, explore endless ideas in one space, edit with brushes, refine layers and elements, and co-create with an AI design agent in real time.
Just think visually and create freely. Because maybe, you were always more creative than you thought!
Stop templating. Start designing now.
Users get 20 free generations per day.
RT + comment in 5hr to get 200 free credits.
10 AI accounts worth following if you actually care about where AI is going:
1. Andrej Karpathy ~ @karpathy
2. François Chollet ~ @fchollet
3. Yann LeCun ~ @ylecun
4. Lilian Weng ~ @lilianweng
5. Demis Hassabis ~ @demishassabis
6. Andrew Ng ~ @AndrewYNg
7. John Carmack ~ @ID_AA_Carmack
8. Fei-Fei Li ~ @drfeifei
9. Jeremy Howard ~ @jeremyphoward
10.Gwern ~ @gwern
Half of AI Twitter is noise.
These people actually build and shape the field.
Who else should be on this list?
New video models coming soon 👀
Seedance 2.1
Seedance 2.0 Mini
Google Veo-4
We're giving away $50 of credits to 10 lucky people who repost, like, comment, and follow @yapper_so!