the way every complex system works is that you deal with problems as they come up. something becomes too onerous to ignore and then you fix it. acceleration & iterative deployment has been the only option: a “pause” in ai development would be entirely squandered

The amount of fraud in education research at elite universities should get way more attention

Few understand how early @amasad was in his predictions and visions
In January 2024, Amjad was talking about how AI would make software creation accessible beyond the traditional programmers
He spoke of a world where you:
- describe in natural language your idea
- the assistant brainstorms with you
- you take a photo of the sketch and they build it
- you share a voice note on the move
- by the time you're home, you have a working prototype
- it built a plan for your review that includes auth, user feedback collection, navigation
All of this now exists
I can't tell you the number of people using the @Replit mobile app to take pics, voice to text, and build their vision on a walk
Humans as the creative heartbeat of the project
It's hard to think back to 2024, but most folks were talking about copilots and assistants to developers
Where AI would be the super productivity tool for engineers and help with code completion
Nobody was talking about the high-autonomy of "build me a whole product"
Since the start Amjad has always wanted to empower everyone, not just the few who have the exquisite knowledge of programming
He has always been saying that AI will become the primary medium between human ideas and working software
The American Dream is in the Cloud
Where anyone can elevate their life through entrepreneurship
https://www.youtube.com/watch?v=kCudFI4tcpg

Grok 4.1 Fast is like a token glitch for OpenClaw 🦞
It's an extremely fast, decently capable model that works GREAT with tool calls
If you want to save money, this is the one. You're probably not going to use it for heavy coding, but for deep research, general tasks and other things, it hits the perfect cost-to-performance sweet spot while offering everything you need
This model is exceptional at search. It performs close to GPT-5.4 (the very next model on the search Arena currently)
The pricing is dirt cheap:
- $0.20/M input | $0.50/M output
- 2M Context window
- Blended ~$0.26 per million tokens
→ 15–25× cheaper than OpenAI GPT or Claude Sonnet ($5–$15+)
Plus, Native Web Search + 𝕏 Search tools are just $5 per 1,000 calls (literally pennies per query)
With a 2M context window, it's an agentic beast for research, support, and deep dives
You don't need to pay 20× more when Grok 4.1 Fast offers near-instant speed + real capability at a fraction of the cost

Friendly reminder, Hinton is the same guy who told us this:
"We should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists.".
That was six year ago.
He was right about the AI. Wrong about the job.
AI already reads scans better than any human. Six years after this nonsense prediction we have more radiologists. That's because Hinton and many others just fundamentally misunderstand that tasks are not jobs and that the job of a radiologist is also interacting with patients, being a light in darkness, providing hope and warmth and care and a thousand other things.
Again we need to stop listening to these folks.
I cannot say it enough, just because this fellow is brilliant with AI does not mean he has any clue how it will impact society. None. Zero.
Rohan Paul: Geoffrey Hinton on AI's job loss:
History’s tech revolutions replaced one job with another. e.g. Tractors replaced farm jobs with factories & office jobs.
But AI will break that cycle, because AI can replace both physical+intellectual labor.

Firecrawl open-sourced "web-agent," a new framework that allows developers to build AI agents capable of autonomously executing "search-scrape-interact" loops on the live web.
By running a single command, developers can scaffold a complete project with streaming UI, API, or library templates.
The framework uses a "Plan-Act" mechanism and can spin up parallel sub-agents to process multiple websites simultaneously.
It is entirely model-agnostic, allowing you to plug in OpenAI, Anthropic (with recent builds defaulting to Claude Opus 4.7), or your own self-hosted models.
Firecrawl: Introducing web-agent, an open framework for building web agents 🔥
Build AI agents that search, scrape, and interact with the web - powered by the same architecture behind our /agent endpoint.
100% open source. Bring any model. Anthropic, OpenAI, or your own.

The countries MAGA describes as collapsing and joyless hold positions 1 through 6 in the global happiness rankings.
They have held them for years.
The country producing the most content about European misery ranks 23rd in happiness.
And 115th in perceived personal freedom.
Not 15th. Not 50th. One hundred and fifteenth.
Six more claims where this came from. All of them fall apart the same way 👇
https://open.substack.com/pub/gandalv/p/the-europe-that-doesnt-exist?utm_campaign=post-expanded-share&utm_medium=web

So @garrytan is ahead of everyone in post AI product development
So much so that even @theo had to agree :)
This is opening a new product design paradigm: AX (Agentic Experience)
Product people need to stop thinking only about UI and UX and start thinking about how to design an AX for their product. This is deeper than you think because it incorporates not only how an Agent experiences your product but how a human user experiences your product through their agent!
I'm coining the term for real here.
Pre-AI, you had to account for what browser and device a user was using, in AX you need to account for what agentic framework or harness is the user using, what core LLM is inside the harness, what communication channel exists between the human and the agent, what tools or skills does the agent have access to, and the list goes on.
Essentially there is a huge portion of product categories that can now be built as "skills" or agent plugins and where a lot of the heavy lifting can be delegated to the agent but as many people have already noticed, an @openclaw user will have a very different experience than a @NousResearch Hermes user using the same product through their agents cause the harnesses are very different!
This is an exciting front for product design and also an opportunity where software and hardware blend because agents operate in both domains!

Re I love Geoff.
But he understands even less than Dario about the effects of technological revolutions on the labor market.
Again, don't listen to AI scientists, as brilliant as they might be, and even less to AI CEOs, as successful as they might be, for questions of labor economics.
Listen to reputable economists who have studied these things like @Ph_Aghion , @DAcemogluMIT , @erikbryn , @amcafee , @davidautor , etc.

Alec Stapp: The amount of fraud in education research at elite universities should get way more attention

my openclaw's been feeling snappier lately. gbrain v0.14 added a shell job type to Minions: ~14 deterministic crons off the gateway, zero LLM tokens.
think: Einstein on cash-register duty at a corner store. calculator handles it now, Einstein goes back to physics.
→ enable the worker (off by default, env flag opt-in)
`GBRAIN_ALLOW_SHELL_JOBS=1 gbrain jobs work`
→ submit a deterministic cron (params as JSON via --params)
`gbrain jobs submit shell --params '{"cmd":"bash scripts/sync.sh","cwd":"/abs"}'`
→ check status (abort flips immediately now, no more 30s stall sweep)
`gbrain jobs list`
Einstein's desk freed up 60%. and when you pull the calculator mid-task, it stops in 2.1s instead of 32s.
Garry Tan: GBrain v0.14 now has a jobs server for OpenClaw that significantly increases throughput and reduces gateway load, based on proven patterns from BullMQ.

Stop taking advice from people who've never built anything. If they haven't put something on the line, their opinion on your risk isn't worth hearing.
The people who judge the attempt are never the ones making one.

The AI to cybersecurity pipeline is real. Every AI enthusiast will eventually become a security practitioner and hacker by default. 🫡
Garry Tan: Sneak Preview: GStack Browser will do defense-in-depth on detecting prompt injections from websites

Do you want an early version of your own Jarvis? Here it is. This repo is incredible and free. Released a few weeks ago and many improvements since then. Created by @garrytan , the CEO of Y Combinator who does this project on the side. Super legit. The project is early and geared for technical users. https://github.com/garrytan/gbrain

Heading to ICLR in Rio 🇧🇷?
We’re hosting our first networking mixer on April 24.
Meet AMI’s technical team and cofounders, and learn more about what we’re building.
Food, drinks, and great conversation included.
Register at https://luma.com/np3x51zh

Yann LeCun was right the entire time. And generative AI might be a dead end.
For the last three years, the entire industry has been obsessed with building bigger LLMs. Trillions of parameters. Billions in compute.
The theory was simple: if you make the model big enough, it will eventually understand how the world works.
Yann LeCun said that was stupid.
He argued that generative AI is fundamentally inefficient.
When an AI predicts the next word, or generates the next pixel, it wastes massive amounts of compute on surface-level details.
It memorizes patterns instead of learning the actual physics of reality.
He proposed a different path: JEPA (Joint-Embedding Predictive Architecture).
Instead of forcing the AI to paint the world pixel by pixel, JEPA forces it to predict abstract concepts. It predicts what happens next in a compressed "thought space."
But for years, JEPA had a fatal flaw.
It suffered from "representation collapse."
Because the AI was allowed to simplify reality, it would cheat. It would simplify everything so much that a dog, a car, and a human all looked identical.
It learned nothing.
To fix it, engineers had to use insanely complex hacks, frozen encoders, and massive compute overheads.
Until today.
Researchers just dropped a paper called "LeWorldModel" (LeWM).
They completely solved the collapse problem.
They replaced the complex engineering hacks with a single, elegant mathematical regularizer.
It forces the AI's internal "thoughts" into a perfect Gaussian distribution.
The AI can no longer cheat. It is forced to understand the physical structure of reality to make its predictions.
The results completely rewrite the economics of AI.
LeWM didn't need a massive, centralized supercomputer.
It has just 15 million parameters.
It trains on a single, standard GPU in a few hours.
Yet it plans 48x faster than massive foundation world models. It intrinsically understands physics. It instantly detects impossible events.
We spent billions trying to force massive server farms to memorize the internet.
Now, a tiny model running locally on a single graphics card is actually learning how the real world works.

Anytime I've built something in Code *without* running it through the gstack pipeline, I've had to come back and rework it. Code is great but it also charges off half-cocked, solving whatever problem it thinks it sees, before asking if that's the *right* problem.
gstack pushes you to take the time up front to ask "what am I actually trying to accomplish, what's the 10x version of that and what'll bite me later if I don't think about it now?"
10-15 minutes that saves hours of rework. Classic measure twice, cut once.
And thank you again, @garrytan!

Re You've been building your Replit project for a while now. You can now finally add a mobile app to it.
5 min instructions below on how to do it safely.
Try it now. $5 credit for updating an existing project, limited time.

one of our investors during YC shared this sales advice on how startups can sell to and close big enterprise companies:
it’s the same method we used at @Crustdata to close a 6-figure deal:
help your champion run an internal hackathon
1) we used to say this: "what if we ran a half-day workshop with your team where they build agents? we would love to sponsor it with our data"
2) let THEM discover the use cases
3) the agents they build become your business case
your champion, just like everyone else, wants to have internal wins to show their boss
imagine running a hackathon with marketers, sales ops, finance people…
and after the hackathon, everyone has built agents using the Crustdata API that help them with their day-to-day work
our champion estimated that the agents built during that hackathon would save them at least 6 figures in vendor costs.
you can now use this number and the success of the hackathon as a case to progress the deal you are working on

This morning, we are launching a seven figure ad campaign against @RoKhanna.
Today is the beginning of the end for @RoKhanna.
We are highlighting Ro's hypocrisy around $600,000,000 of trades since being elected to Congress.
You'll find it on Hulu, YouTube, Instagram, and eventually linear.
If you live in the district, you'll also be getting a mailer about the contrast of our policies.
We're in this to win it.

1. Go to an existing Replit project
2. “Make a mobile app”
3. Publish to App Store
Zhen Li: Now you can add a mobile app to the website you’ve been building on Replit.
Ship it to the App Store from your existing project.
Just open your app and ask the Agent.

New post w/ random thoughts on AI (thread)
I will probably get a # wrong, but here we go :)
1/12 OpenAI & Anthropic now at 0.1% of US GDP *each*
In a year, AI revenue likely to be 1-2% of US GDP
What does AI mean for US GDP growth? Does productivity get lost mismeasured a la internet era?

.@TomSteyer got rich building ICE prisons like this one. Now he’s using that fortune to try to buy this election. Californians deserve better.

You can now add a mobile app, slide deck, or animation to any web app you have built prior to Replit Agent 4.
Everything lives in the same project, with shared branding & data.
Just open your app and ask the Agent what you want to build. You also get $5 in credits to get started for a limited time
Zhen Li: Now you can add a mobile app to the website you’ve been building on Replit.
Ship it to the App Store from your existing project.
Just open your app and ask the Agent.

I’m excited to lead HyperWrite’s next chapter. I appreciate the trust Matt, Jason, and our investors have placed in me.
For those I haven’t met yet: I’ve been with HyperWrite since the beginning and have led development on many of the key features we’ve shipped.
As CEO, I’m focusing HyperWrite on in-context writing, starting with TypeAhead in the Chrome extension: personalized predictive writing that carries your context across ChatGPT, Gmail, Google Docs, and more.
Try the Chrome extension. Link in first reply.
Matt Shumer: Quick update I’ve been meaning to share for a while: @JasonKuperberg and I stepped away from HyperWrite a few months ago. 
We’re proud to have built one of the first consumer AI products, and turned it into a profitable business with real scale.
But the AI landscape is moving

Can confirm. Fiancée spends weeknights scrolling puppies on OLX (ads, pagination, page 1 of 100).
I opened @Replit, typed: "puppy listings → Pinterest feed → pinch zoom."
Came back from coffee to a live app with API.
AI-generated app. Puppies are real 🐶
Check for yourself https://realpuppies.dog/
Amjad Masad: 1. Go to an existing Replit project
2. “Make a mobile app”
3. Publish to App Store

The orchestration logic in @garrytan’s gstack is underrated.
The built-in compacting phase before every plan execution is the real win, it forces a project state sync at the start of every session, effectively killing context rot in long horizon agentic workflows.

I don't need any of it for my work (or maybe I do), but I need to upgrade to a pro plan on Replit now. @amasad is cooking amazing stuff!
This Lego bricks simulator was the first thing I made when I tried Replit, and was amazed at how seamless it was.
Amjad Masad: 1. Go to an existing Replit project
2. “Make a mobile app”
3. Publish to App Store

AI expert @mattshumer_ discusses his most recent contribution to NOTUS Perspectives, on the intersection of moral authority in Washington and artificial intelligence regulation.
https://www.notus.org/perspectives/where-should-washington-look-for-moral-authority-in-2026#matt-shumer

been using gbrain for a while and never actually opened the soul-audit skill(for your openclaw/hermes). ran it today. came out better than i expected.
it's a 6-phase interview that generates a fresh agent, every answer is from you:
→ phase 1 identity: what is this agent to you (SOUL)
→ phase 2 vibe: pick one of formal / direct / technical / casual (SOUL)
→ phase 3 mission: top 3-5 things you want it to do (SOUL)
→ phase 4 who you are: role, projects, key people (USER)
→ phase 5 boundaries: who can see which layer of your brain (ACCESS_POLICY)
→ phase 6 cadence: when the agent pings you (HEARTBEAT)
the agent it built knows me way better than the ones i usually throw together. templates in the repo are clean too, so if you want to re-tune one section later you flip the template.

🤖 This debate is largely a matter of productivity gains-@a_bergeaud's latest book is a must read to easily understand why !
📈 With @BunelSimon @XJaravel @tmikaelsen @alexandraroulet J.Søgaard we found a rise in employment among early AI adopters, even for substitutable jobs !
Yann LeCun: Dario is wrong.
He knows absolutely nothing about the effects of technological revolutions on the labor market.
Don't listen to him, Sam, Yoshua, Geoff, or me on this topic.
Listen to economists who have spent their career studying this, like @Ph_Aghion , @erikbryn ,

AI's potential negative impact on democracy is discussed ad nauseum, but what about the upside of AI-powered civic tech?
Check out @garrytan engaging with Garry's List subs to crowdsource and mobilize a movement. Every mayor and governor should copy this playbook. The best will.
They'll reap the rewards as their constituents experience engagement from their officials that was impossible pre-AI
Why can't we codify our constitutions, like a claude.md for government, enabling AI watchdogs to automatically flag when politicians overstep their authority?
What would that look like applied to budget management and fiscal responsibility?
There are probably infinite other ways that AI can increase accountability and effectiveness in a modern democracy. The faster these get proven out, the less likely we are to throttle our potential out of fear

Totally with Garry here.
I'm 58 and still building stuff at 2am.
I was similarly obsessed when building for the early web in the 1990s. But this is way more fun.
Garry Tan: I came back to code because AI made it possible for me to build at a level I couldn't before.
I'm not coding despite being CEO of YC. I'm coding because this is the most important technological shift since the internet and I'd be an idiot to experience it from the bleachers.

some ideas are much clearer when you can use coding agents to show a proof of concept.
eg I hadn’t really understood how GPUs vs NPUs compete for memory on devices until this writeup (w/ its accompanying code)
Vivek Gupta: http://x.com/i/article/2046347462399655936

Earlier this year Yann LeCun left Meta because Mark Zuckerberg wouldn't bet the company on JEPA. Last week his group dropped the first JEPA that actually trains end-to-end from raw pixels. 15 million parameters. Single GPU. A few hours.
The timing is not a coincidence.
For four years Meta has been the house that JEPA built. LeCun published the original paper from FAIR in 2022. I-JEPA and V-JEPA came out of his lab. The architecture was supposed to be the escape hatch from LLMs, the path to robots that actually learn physics instead of hallucinating about it. Every version shipped fragile. Stop-gradients. Exponential moving averages. Frozen pretrained encoders. Six or seven loss terms that had to be hand-tuned or the model collapsed into garbage representations.
Meta kept funding LLMs. Llama shipped. Llama scaled. Llama got beat by Qwen and DeepSeek. Zuck spent $14 billion to buy ScaleAI and install Alexandr Wang. The FAIR robotics group was dissolved. LeCun's research kept winning papers and losing the product roadmap.
He left, started AMI Labs, and said publicly that LLMs were a dead end.
Now the paper. LeWorldModel. One regularizer replaces the entire pile of heuristics. Project the latent embeddings onto random directions, run a normality test, penalize deviation from Gaussian. The model cannot collapse because collapsed embeddings fail the test by construction. Hyperparameter search went from O(n^6) polynomial to O(log n) logarithmic. Six tunable knobs became one.
The downstream numbers are what should scare the robotics capex class. 200 times fewer tokens per observation than DINO-WM. Planning time drops from 47 seconds to 0.98 seconds per cycle. 48x faster at matching or beating foundation-model performance on Push-T and 3D cube control. The latent space probes cleanly for agent position, block velocity, end-effector pose. It correctly flags physically impossible events as surprising. It learned physics without being told physics existed.
Figure AI is valued at $39 billion. Tesla Optimus is mass-producing. World Labs raised $230 million to sell generative world models. Everyone in humanoid robotics is burning capital on foundation-model pipelines that plan in 47 seconds per cycle.
LeCun's group just showed you can do it with 15 million parameters on a single GPU in a few hours.
This is the Xerox PARC pattern running again. Meta had the next architecture. Meta had the scientist. Meta dissolved the robotics team, passed on the productization, and watched the exit. Three months later the lab that was supposed to be Meta's publishes the result that resets the robotics cost structure.
The paper is worth more than Alexandr Wang.

It's so funny to look back at how things start... :) Go @openclaw !
Peter Steinberger 🦞: warelay was meant as simple app that triggers a cli
now it's a multi-agent orchestration system and I'm about to add Telegram.
Send help... or a better name.