Michele Catasta
@pirroh
Followers
11K
Following
4K
Media
112
Statuses
1K
President @Replit | Head of Applied Research @Google | Research Scientist + Instructor @StanfordAILab | 🇮🇹 @🇺🇸
Joined February 2009
I can't tell you the last time I was so excited to see a new AI capability in action. We plugged in Claude computer use in @Replit Agent as a human feedback replacement. And… it just works! I feel it won't take long until our agent will become fully autonomous.
30
125
1K
🧑💻 replit-code-v1-3b is out!. Head to our HuggingFace 🤗 org page:.to use the open-source release of our ReplitLM specialized on code completion. This will be the first of many LLMs 🚀.1/ 🧵.
22
180
897
Announcing Replit Agent v2, available in Early Access today!. More highlights & how to get started in 🧵
47
82
816
Sundar is our Ideal Customer Profile. PM turned builder… with a long detour as CEO of a $2T company 💪 .Welcome to @Replit, Sundar!.
"I was Vibe Coding with Replit a few weeks ago. I mean the power of what you're going to be able to create on the web; we haven't given that power to developers in 25 years" — Sundar Pichai
23
35
789
You will notice the difference immediately. Better performance, straight to the point. We've been optimizing Replit Agent for Claude Sonnet 4.0 during the past weeks. The upgrade has been rolling out to our users since early this morning -- you will all get it by end of day!
66
39
472
US immigration is deeply flawed. It took me years (and several visa permits) before I could stop being anxious about my residency status. Lots of people helped me in the process. Kudos to @0xSigil for creating a network of people who went through the same struggle. Now it’s our.
14
16
397
I began working on Transformers in 2018 at @StanfordAILab -- training and inference in the early days was anything but fun. Today it takes 3 lines of code and zero setup time to use a multi-B LLM on @Replit. Releasing Replit ModelFarm is one of my career dreams come true. 1/ 🧵
7
40
371
Want to create your own 3D game? Never been so easy. Try our brand new Game Stack, now available on Replit Agent!. This FPS game 👇 took me only 4 prompts.
18
41
349
While developing @Replit Agent, we constantly face the challenge of how to interact reliably with a large number of external services. Similarly to what Web standards enabled in the past 30 years, we need a lingua franca between AI agents and the Web ecosystem. MCP has.
Introducing the Model Context Protocol (MCP). An open standard we've been working on at Anthropic that solves a core challenge with LLM apps - connecting them to your data. No more building custom integrations for every data source. MCP provides one protocol to connect them all:
4
23
348
Once again, @AnthropicAI pushed the boundaries of LLM coding capabilities. We loved working with a preview of Sonnet 3.7 during the past weeks -- expect an announcement about Replit Agent very, very soon!
6
16
337
It’s that good — @AnthropicAI cooked once again. First model of this class with frontier CodeGen capabilities. ↓ Claude 3.5 Haiku soon in @Replit agent!
6
21
318
Be Precise. Be Well-Organized. Be Patient. Coding agents have a much gentler learning curve compared to software development tools. Still, learning a few skills is necessary to be productive with them. Start from this video ↓
8
28
316
We just put a coding agent in everyone's pocket. We chose to do it not because it was easy, but because it was hard. One giant leap towards empowering the next billion software creators.
Whatever you need… make an app for that. Now on your phone. For everyone. Free.
12
21
298
If you are familiar with everything covered in this presentation, I got a job for you. Join me and the AI team at @Replit to build the most impactful project of your career — expect intensity, intellectual challenge and outsized glory once the mission is accomplished!.
Curious about how Replit Agent works? At @Replit we hosted tech talks discussing agent's internals. Let's dive into what goes into an LLM that powers the agent - slides from my talk. 🧵
5
24
193
Soon we'll ship @yoheinakajima's workflow ↓ as a native feature on Replit Agent, with a very intuitive UX. Test-time compute can be scaled in so many effective ways, not only for thinking tokens!.
I now regularly have multiple @Replit agents tackling the same project with different approaches, and then when one works better than others, I fork it into three new approaches…. Kinda feels like one of those maze algorithms if you know what I mean.
3
7
163
LLM hyperscalers get all the "Attention" -- it's unarguable how capable their models are. Which begs the question: is it even worth to train bespoke LLMs on a limited budget?. My answer is yes, especially after months of using our Code LLM in production at @Replit. 1/ 🧵
7
26
149
When we say ASAP, we usually mean "shipping in a few hours" 🚢
We loved being early customers of the new Claude 3.5 Sonnet model, and we'll make it available on @Replit Agent ASAP!. Learn more about this amazing release by @AnthropicAI:.- -
6
12
148
~200 AI builders at @SHACK15sf for the @AnthropicAI Claude 2 Hackathon. Lots of interesting projects and familiar faces in the AI space!. Here are my highlights from judging the final round 👇.1/ 🧵
3
18
143
The groundswell of interest for Llama-1 (and for all our previous open source AI packages, Like PyTorch, DINO, SAM, NLLB, wav2vec. ) is what convinced the Meta leadership that the benefits of an open release of Llama-2 would overwhelmingly outweigh the risks and transform the.
1
3
110
"We have not yet discovered the optimal UI paradigm for generating software using AI. It’s going to take determined, open minded people to discover it.". Jordan is bringing his determination and outstanding expertise to Replit. I couldn't be more excited to work with him!.
Exciting News: I am joining @Replit to lead product development!. A couple months ago, I started using Replit for my own AI driven coding projects and it was immediately clear that AI is going to enable the next billion programmers. And Replit made it so easy. I have decades of.
6
7
107
Having worked for both teams involved in this partnership, I’m in the privileged position of congratulating both my former colleagues @googlecloud and current ones @Replit. Today we are one leap closer to a future with 1B developers in the 🌎 🌍 .
2
6
103
Yesterday we hosted @jordwalke (creator of @reactjs) at the @Replit HQ for a fireside chat with @amasad. Jordan has been using Replit Agent a lot lately, so he asked us “how long until LLMs will be 2X better than today?”. According to 👇 not only LLMs will improve, but the next
9
15
93
The frontier of coding agents keeps expanding rapidly👇 .Working at Replit feels like being an explorer, always searching for new capabilities. If this is your passion, send me a DM — I’m always searching for the best AI engineers!.
The strongest capability unlocked by @AnthropicAI Sonnet 4 is excellent multi-tool calling—it lets an agent do 10× more work in a single response and reach its goal 10× faster!
5
8
92
🚀 .@AnthropicAI is providing the kerosene, @Replit is building the rocket. Full article by @Inc in the comments.
4
8
92
While building @Replit Agent, @hwchase17 and the whole @LangChainAI team have been awesome partners. Read my interview on the "Breakout Agentic Apps" series↓
2
34
91
Claude Opus 4 and Sonnet 4 push the envelope once again on coding capabilities, handling complex multi-file changes without touching code you didn't ask to modify. The improved precision in instruction following plus thinking mode gains show real potential to fundamentally.
Introducing the next generation: Claude Opus 4 and Claude Sonnet 4. Claude Opus 4 is our most powerful model yet, and the world’s best coding model. Claude Sonnet 4 is a significant upgrade from its predecessor, delivering superior coding and reasoning.
5
1
91
We loved being early customers of the new Claude 3.5 Sonnet model, and we'll make it available on @Replit Agent ASAP!. Learn more about this amazing release by @AnthropicAI:.- -
5
2
85
"Zero to One isn't cool. You know what's cool? Scaling." -- improbable concoction of Sean Parker and Peter Thiel. 2022 was the 0-to-1 year for AI at @Replit. In 2023, we scaled up AI in many dimensions: features, models, data, team, etc. Time for a walk down memory lane 🧵
9
12
87
Don't waste time debating if AI Agents are all hype or not. They are already here, in production ↓↓↓."Basis is deployed today, taking on real work, and growing fast.". The job market will change completely, starting from next year. Get to work on agents today. You can thank me.
We’ve raised $34m to accelerate our goal of equipping every accountant with a team of AI agents. Led by Keith Rabois at Khosla Ventures with Nat Friedman & Daniel Gross and continued participation from Better Tomorrow Ventures, BoxGroup, and Avid Ventures. Excited to also
5
3
88
No matter how you are building on Replit, you can always use the best models from @AnthropicAI. Claude 3.7 Sonnet, now default on Assistant.
4
3
82
7
0
75
Realtime app design preview.You can provide feedback before moving forward with your build.
1
7
72
Fundamentally, the recipe for reliable agents is: high intelligence in a guardrailed environment. While LLMs keep improving, we are creating an ecosystem where Replit Agent thrives. DB and Auth are the first milestones, but more will come. The compound effect has just started.
@Replit Auth is now available to all. Ask the Agent to "add Replit Auth authentication," and give your app a login experience with a plethora of social providers, email/password/forgot password, user management, and email lists all with zero setup. Let me know how it is!
4
5
72
Agent v2 is now enabled for everyone. Thanks for all your feedback during the Early Access phase — it pushed us to ship something we are deeply proud of. Keep it coming!.
Testing @Replit Agent v2 for over a month—metrics have been mind-blowing. v2 uses reasoning to be 10x more autonomous. Insane to watch do 15-minute debugging to recover from gnarly bugs. Don't take my word, try it: Free users get 20 checkpoints and Core get $10 (limited offer)
10
5
70
Agent v2 is Mckay-approved: achievement unlocked.
@pirroh Tried a super simple prompt and got a terrific UI with all pages working + working Postgres DB. I hit send and like 5min later was done. Great experience - v2 is awesome
5
3
70
Tune in for the opening keynote at @aiDotEngineer Summit, today 3:15pm PT. @amasad and I will announce a big AI launch 🚀 on @Replit … as well as a small (but mighty 💪) launch!
2
4
67
Replit Agent wouldn’t have been possible without @braintrustdata — IYKYK.
We’re thrilled to announce that we've raised a $36M Series A led by @martin_casado at @a16z to advance the future of AI software engineering, bringing our total funding to $45 million. We’re also introducing functions — a flexible primitive for building with foundation models.
1
4
66
Making @Replit Agent better day after day requires a lot of exploratory work -- synthetic data generation? model finetuning? You name it, we did it. Thanks for the shoutout during @Meta Connect, @_chriscox! .Replit AI ❤️ 🦙
4
49
59
Impatient to try it? Check out our blazing-fast ⚡️ GPU-hosted demo, courtesy of the amazing folks at @huggingface:.3/.
3
7
55
If you want to learn more about Agents from a research POV, check CS329A: @achowdhery and @Azaliamirh are doing an amazing job with this course, and I'm grateful they brought me back to teach at @Stanford after a 5-year hiatus 🙏.
1
5
53
As always, the special ingredient is in the training data -- a careful mixture of real-world errors (collected on Replit) combined with synthetically-generated code fixes. 4/6
2
3
52
One step closer to the “everything app” for creation. @Replit 🧡📱.
Announcing Native Mobile App support on Replit. Now you can build iOS and Android apps that you can take all the way to the App Store without writing any code, powered by Replit Assistant. This is early access—full Agent support coming soon!
3
2
49
$$$$ development team >> $ Replit Agent.
People who complain about the cost for credits on tools like @Replit have clearly not paid for a development team before :).
7
5
51
Ask for any 3D model you can imagine, and Replit Agent will generate it for you right away.
6
3
50
You’ll be working closely with me, so I’ll throw in also a few extra objective functions:. - Italian accent 🇮🇹 .- hand gestures 🤌.- pasta cooking 🍝 .Multimodality FTW!. Apply on 👇 this 🧵.
If you're a great ML engineer dying to get into LLMs but you have no experience, we'd happily take you and train you. Come train code models to be used by millions of people!.
1
3
48
It takes a builder to understand how unique Replit Agent is. Joe's thread is packed with pearls of wisdom ↓.
After a month of building my SaaS project with Vibe Coding and @Replit, I'm blown away by the experience! As someone who's launched dozens of products over 15 years as a Sr. Product Manager, I've never felt this confident about security. Here's why! .@amasad .@pirroh .@mattppal
2
2
49
Come chat about agents with @hwchase17 and yours truly!.I’m always down to meet AI Engineers that would love to make @Replit Agent even more powerful — one more reason to attend if you are in town for the OpenAI Dev Day 🌁.
How did @Replit build an agent that creates fully functional apps in just minutes? Get the inside scoop with President of Replit, Michele Catasta (@pirroh) at our meetup on Monday, Sep 30th in San Francisco. 🤓🪄. Join Michele and our CEO, Harrison Chase (@hwchase17) for a
6
8
49
It's getting easier by the day to fine tune LLMs! For example, QLora on a 65B LLM with a single 48GB GPU: . Training LLMs from scratch like we did at @Replit.is certainly more work, but companies like @MosaicML and @huggingface make it way more accessible.
How difficult is it for small organizations to tune/build their own LLMs? Will every company eventually have its own little datacenter to train these LLMs? @husseinmuhaisen @GabiImmelman.
3
5
42
Brand new app creation experience.We guide you throughout the entire ideation process, recommending you which steps to take.
1
2
46
How do we deploy LLMs for millions of @Replit users with such a small team?.We hire the *best* engineers, that’s how!. Check @nurblieh cloud wizardry 👇.And then… /1 🧵.
Replit AI engineer (and YouTube’s first SRE!) @nurblieh on how we serve LLMs super cheaply using preemptible GPUs and how we cut cold start times by an order of magnitude
2
4
43
For all the details, check our technical report:. Huge congrats to @madhavsinghal_ Ryan Carelli @giansegato @vaibhavk97 and the whole AI team 👏. 6/6.
6
3
43
TIL Smiling Curves might be a good model also for the AI value chain. Great talk @swyx 🙌.Link to the slide deck ⬇️
0
6
43
Spoiler: “Replit is #1”. And @rileybrown_ai wasn’t even using Replit Agent v2.
Have you ever vibe coded 6 versions of same app at the same time using all of the best AI coding tools. Well I did it in this video:. Tools used: .Cursor @cursor_ai .Windsurf @windsurf_ai .Replit @Replit .Bolt @boltdotnew .Lovable @lovable_dev .v0 . TimeStamps.00:00 -
5
3
42
Eager to make other Fermi estimates on LLMs?. Check out one of my favorite resources, carefully curated by @anyscalecompute (@robertnishihara @waleedk) ↓. "Numbers every LLM Developer should know": 4/.
2
2
40
Mayur and I share the same research passion. Public endorsements make me even more proud of the work we are doing, especially when they come from a renowned professor in my field 👇
This weekend I gave @Replit's AI agent a try with my 12 yo who has no formal training in computer science. In ten minutes each, he built and deployed two working apps. The number of CS technologies involved is mind-boggling:.- web dev (Flask/React).- application logic (Python/JS)
2
2
39
Buy yet another device (for $169!) or turn your smartphone into an AI-powered meeting transcriber with a custom Replit App?. personal apps >> multiple personal devices.
Built an Ai powered meeting transcription app in @Replit today. I was debating buying Plaud but thought about it & built my own. The mic stays on even if you lock the screen so it effectively records your meeting, transcribes it then gives you Ai feedback for action items,
1
4
38
LLMs are all the craze, but only few AI specialists ever had a chance to build them from scratch. If you want to understand the whole lifecycle of LLMs (down to the nuts and bolts), this is your dream job. Come work with us at @Replit!.
3
5
38
Before I verify the Scaling Laws on my ever-growing white hair, join our LLM team 👇.
Replit is 85 person team and while we are proud of our impact relative to our size, we really need some help. Particularly on platform engineering & AI/ML.
1
3
35
We took a 7B Code LLM and finetuned it into a tool that mimics the behavior of LSP Code Actions. 3/6
2
2
37
Python has one of the most advanced software toolchains, yet only ~10% of the errors are automatically repaired by LSPs. We sifted through the hundreds of millions of diagnostics trigger on Replit every day, and realized that more than half of them could be fixed with AI. 2/6
1
3
36
We run evaluations based on academic benchmarks, as well as on a collection of real-world diagnostics on Replit. The results speak for themselves -- we are on par or better than vastly bigger LLMs, at a fraction of the latency and cost. 5/6
2
2
36
LLM hacker willing to help us?.Send us PRs, and apply right after for an AI position at Replit: 7/.
3
3
35
Banger‼️ .Also, this is one of the many pitfalls that Replit dodges for you.
Introducing VibeCon — the world’s largest vibe coding conference. Make sure you register today: http://127.0.0.1:8080/register.
2
2
35
Ambitious companies don’t shy away from hard problems. We built our own Copy-On-Write distributed file system — before LLMs became so good at coding, and with a much smaller team than DeepSeek. Soon Replit Agent will have unique features based on it. You can just do things.
looking at the open source repositories from deepseek doesn't make any sense to me. like, for their relatively small company size. they wrote their own filesystem for gods sake. but then i realize, that they're ai augmented.this is what happens when you use AI to make software.
2
2
34
spot on, that's exactly the balance we are trying to strike with Replit Agent. Aristotle said: "in medio stat virtus" (virtue is in the middle).
closer to the code = more powerful, flexible, slower to start. more abstracted = quicker to start, slower for more complex building. replit is best of both worlds imo
1
3
34
“People who are really serious about software should make their own hardware.” [1982]. More than 40 years later, Alan Kay’s quote still sounds as relevant as it gets – no wonder he is recognized as one of the most influential pioneers of our industry. As a matter of fact, over
1
3
29
I cherish a lot of memories, but remembering the crontab(5) syntax is not one of them. Just tell Replit AI the schedule you have in mind, and we’ll take care of everything else!.
Today, we’re launching Scheduled Deployments. Build your app, then tell Replit to “Run this code every Tuesday and Thursday at 3:00 PM.” We’ll handle the rest. Spend less time configuring and more time shipping.
0
0
30
When it comes to LLMs, it's easy to build products, but not easy to maintain them -- progress is so fast-paced that it's hard to predict if what works today will still be the best approach by end of year. Still, In-Context Learning + Retrieval work great for now! Let's build 🚀.
It seems that retrieval and context management are all the rage now. Is that premature optimization? Do you think there is an easy buy vs build decision there? @sidravi_.
3
3
26
On the last episode of @CogRev_Podcast , @labenz asked me a long list of questions about @Replit and our AI roadmap -- I was impressed by how deeply Nathan understands our mission. I just went with the flow, and we ended up talking AI for almost 2 hours!. Links below ↓.1/ 🧵
1
5
27
They will keep being powered by Neon — expect them to get even better over time 🤝 @nikitabase @matei_zaharia @alighodsi.
what happens to replit databases that were powered by neon?.
1
0
28
Heavily subsidizing frontier models promotes unsustainable business models, where only AI model providers and hyperscalers would thrive eventually. This time around, we can prevent another tech era fueled by ads. I’m grateful more product builders are being upfront about it 🙏.
If you're shocked at the price of using @cline with frontier models, understand that this is the actual price of inference. It's not realistic to provide full-powered AI at $15-20/month without severely limiting what these models can do.
3
3
28
"Small" Language Models are the sweet spot that will bring #AI4Code to every developer in the world. Low latency is paramount when it comes to UX!. @amasad, @samipddd and the whole @Replit team accomplished the unthinkable today. I couldn't be more proud.
In 2018 we announced Multiplayer Mode. Since then, Replit became the world's leading collaborative editor. Today, we're announcing AI Mode, which infuses state-of-the-art intelligence into nearly all IDE features, the next major evolution for Replit:
2
3
27
A new quote for the Replit yearbook. Our company history, since 2016, is one of steady and tangible growth rather than of chasing the hype – it’s deeply rooted in our culture. We can see the road ahead that will lead us to ADI, and we are paving the way for it – line of code by.
3
1
24
TIL: Wikipedia does *NOT* have an article on Graph Neural Networks 🤯. Call to action! /cc.@m_deff @thomaskipf @williamleif @PetarV_93 @PeterWBattaglia @mmbronstein @weihua916 @rusty1s @chamii22 @guennemann @marinkazitnik @jure @HazyResearch.
0
7
27
