Ryan Carson
@ryancarson ·
I'm live now - stop on by
Back
AI posts
AI posts
People and companies shaping AI. Filter your watched accounts, catch up on recent posts, or find the most liked and viewed.
Compiled by the Toolstash team · Updated
209 watched accounts · 2978 matching posts · Showing up to 3,000 recent stored posts. Counts are snapshots, not live X totals. This page refreshes every minute while visible. Last collected .
Alex Lieberman@businessbarista ·
Micro1 is a $4B company run by a 25-year-old, and it sits in the middle of the fastest-growing corner of AI: the data business.
I sat down with @aliansarinik, micro1's CEO, to break down how AI models actually get smarter and who gets paid to make it happen.
Complete list of notes (and link to interview below)...
Big takeaways: - Data labeling is a $30–50B/year market. Ali thinks it hits ~$1 trillion a year within 2–4 years, comparable to compute spend. - The internet was a one-time gift. Pre-training ran on decades of human-created data. Now that it’s used up, every frontier gain requires net-new data created by human experts. - micro1 doesn’t sell data. It sells “units of intelligence improvement,” meaning measurable gains in a customer’s model or agent. - Synthetic data isn’t the bear case. Ali calls it the most bullish case for his business.
How micro1 got here: - It started as a dev agency Ali ran at Berkeley. His bottleneck was interviewing engineers, so he built an AI screener. - That screener became one of the first AI recruiters. People hated on it, until it became normal. - After scaling to a few million in run rate as a talent marketplace, they saw the data industry shifting toward expert data. The hard part of expert data is recruiting experts, which was exactly what they’d built.
Why spend keeps going up: - Model progress = compute + algorithms + data. Ali argues data is what actually pushes intelligence forward. - Post-training needs experts distilling their judgment into RL environments. It’s expensive, but the ROI is there. - The trillions in compute bets assume models keep unlocking new capabilities. If they don’t, those investments are a disaster. So labs have no choice but to keep buying data.
How a data pipeline actually works: - Step 1: Find the failure. Either a lab hands over a spec, or micro1 diagnoses where the model breaks and pitches the pipeline. That research arm is increasingly important. - Step 2: Write the spec. Domain, sub-domain, data type, and volume (usually hundreds to tens of thousands of data points over 1–3 months). - Step 3: Recruit hyper-specific experts. Never “lawyers.” More like M&A lawyers who are great at writing legal memos. - Step 4: Turn expertise into training signal. Experts create prompts the models fail on, then write the rubrics and verifiers that define a good outcome. AI models assist them throughout. - Step 5: Manage the pipeline like a factory. Track data quality, throughput, and whether the customer’s model is improving without regressing elsewhere.
Two ways experts grade models: - Likert-scale evals: experts rate outputs 1–5 on dimensions like completeness or verbosity. Great for subjective, pre-release signal. Not reusable, because every new model means bringing experts back. - Rubrics: experts write 20–40 binary criteria for a great answer (e.g. “valuation range must fall between X and Y”). More work up front, but a grader model can re-score every future model with no new human time. This is what powers actual RL training runs.
The hard parts about this business: - Vetting at scale. micro1 interviews tens of thousands of people a day, with 10M+ through the process. Doing that manually would take thousands of recruiters. - Managing experts. Strategic project leads run each pipeline, hold office hours, and keep experts aligned. - Quality in subjective domains. Most fields are more subjective than people think, so they need peer review (lawyers reviewing lawyers) plus AI-based QA. - Reward hacking. They recently caught a Gemini model searching its sandbox for hidden files containing answers. If cheating gets rewarded, models learn that cheating works and become misaligned.
The enterprise opportunity: - Ali’s bet: Fortune 500/1000 companies converge on ~$100M/year in data spend. That’s roughly 1/10th of a lab, but there are 1000x more enterprises. - It’s not a software budget. It looks more like payroll: core IP. Framed that way, it’s 1–5% of payroll. - “Owning your own intelligence” means defining exactly what you want agents to be good at and building an eval stack to measure it. With that stack, models are swappable, open or closed. - Most enterprises can’t actually use their own data. It’s customer data, so they can’t train or even evaluate on it.
Why everyone’s buying SMB data: - An RL environment has three parts: seeding data, tasks, and tooling. - Ali calls it “scaling on realism,” almost a new scaling law. The more realistic the environment, the better the downstream model performs. - So micro1 licenses operational data (Slack, Google Docs, etc.) from SMBs, transforms the PII instead of redacting it to keep the noise intact, and seeds environments that act as digital twins of real workplaces.
Why synthetic data makes the pie bigger: - Today, a lawyer writes the memo and the rubric. Soon, the lawyer writes the memo and a model generates the rubric, cutting expert time 50–60%. - Extrapolate: the model does ~99% of the work and the expert signs off. A 15-hour task becomes 5 minutes. - When environments are that cheap, pipelines grow 1000x. Experts become 0.1% of the work, but demand for them still goes up. Jevons paradox. - It’s already happening in coding: 70–80% synthetic, and micro1 still can’t hire coders fast enough. - Margins shift from expert hours to compute, which is far more controllable. - Ali’s framing: if you could buy IQ points for your systems almost free, you’d buy all of them.
The “infinite last mile”: - Knowledge isn’t static. Ali cites a stat that a big share of what med students learn needs relearning within a few years. - AI accelerates that churn, which creates a feedback loop. There’s no finish line for aligning models with humans. - Labs today concentrate spend in four domains: medicine, finance, legal, and coding. Every other domain is still waiting on cheaper data to unlock.
Which corner of the AI stack should I break down next?
Link to the full convo with Ali: https://youtu.be/2vL20qxD0BM
Joshua Saxe@joshua_saxe ·
Congrats to @AlexTamkin et al; surgical research design and concerning (if intuitive) results showing just how readily experts adopt AI agents' interpretations of evidence (reasonable to expect this'll increase with AI capabilities). @ezraklein I hope your interviews on AI safety and societal impact scope this in
Riley Brown@rileybrown ·
Yeah keep in mind this is all for work.
I want to agent max in my professional life but tbh for my personal life I want to read, write and play poker & get the fuck off my phone.
The main problem - agents have made me so productive, and given me so much leverage (especially with content) that it’s hard to stop working.
elvis@omarsar0 ·
Decision models can also improve your OCR pipelines.
You can also use Jev-like models for verification.
Here is a great example of this:
MiniCPM-OCR from OpenBMB and ModelBest pairs a small open document parser with a second model that scores the reliability of the OCR output.
The parser has only 1.3B parameters. It handles Chinese and English documents, including scans, photos, curved pages, and tilted inputs, and it supports key information extraction.
The second model, MiniCPM-OCR Jev, checks extracted text, formulas, and tables. It also works on outputs from other OCR models, so you can run it on top of the parser you already use.
More details below.
Agents now read PDFs, invoices, and papers all day. One bad table extraction carries into every step that follows.
A simple setup looks like this.
> Parse the document with MiniCPM-OCR > Score the extracted text, formulas, and tables with Jev > Send low-scoring output to a human or a correction step
Jev estimates reliability. It doesn't fix errors itself.
The weights are open, with a fine-tuning framework and inference support for llama.cpp, vLLM, SGLang, and Ollama. There are also Agent Skills integrations.
If you build document agents, test the parser and Jev together on your hardest pages, like dense tables, formulas, and phone photos.
jack friks@jackfriks ·
my wife is now working on my couples app with me
can i now say i run a family business?
@levelsio@levelsio ·
Anyone know a good organic free roaming chicken supplier in Portugal?
Turns out my organic free roaming butcher was selling me industrial chicken 😂😂😂😂😂
- Ryan Carson
@ryancarson ·
Secrets to become pro agent manager! https://x.com/i/broadcasts/1lJQRWOlVVvxE
Elon Musk@elonmusk ·
At Tesla, we even care about the chipmunks
dax@thdxr ·
lot of people making stuff up so here's some more context
you can imagine there's hidden enterprise demand for open source models that isn't visible
but if you look at the top inference providers that aggressively go after this business, they have shared some numbers and they are all roughly the same size as us. and at best there's 5 of them
if you think enterprises are self hosting inference on their own GPUs, there are so few people who can run reliable inference and they get paid a lot and work at ai labs. and if they really want sovereignty they're not getting that by leasing GPUs from a third party
the root point here is this is entirely the wrong way to look at things, you just have to look at products
how many products exist with a lot of traction that can use open models? this is what creates demand, it's not some magic thing that grows on its own
OpenCode is one of those and we do see a lot of enterprise inbound. but it's so early, all of these companies are just starting to look at mixing in some OSS inference
so yes demand will probably go up in the future but everyone in this industry is investing resources in preparing for that future. and so few people are building the products that will make it happen
Romain Huet@romainhuet ·
Bonjour, Paris! 🇫🇷
Always special to be back home. Full house for DevDay Exchange, and we’re about to get started.
Can’t wait to reconnect with the French developer community and see what you’re building!
- @levelsio
@levelsio ·
Product Hunt permissions to Login with X went absolutely mental
Safer to not even use it anymore then?
Simon Willison@simonw ·
I think October represents another inflection point in capabilities for coding agents and the models that drive them: Claude Opus 5.5 can now build video games that are actually fun https://simonwillison.net/2026/Oct/28/vibe-game-development/
Daniel San@dani_avila7 ·
Payments from the CLI? I like where this is going.
Well… I can already order Chipotle with the DoorDash CLI
Being able to send payments with Stripe feels like the next logical step
Mistral AI@MistralAI ·
Today, we’re introducing the Responses API in public preview, made for building agents on Mistral.
Each response returns the model’s answer, its reasoning and its tool calls as separate items, so your app can handle every step more easily than with Chat Completions.
It works with all chat models hosted by Mistral, including Large 4.
Matt Pocock@mattpocockuk ·
Debating just straight-up banning useState in my React apps
Push the agent into the black hole of success
- dax
@thdxr ·
with supply chain attacks becoming a bigger problem AUR really has to become more sophisticated or get replaced
it's impossible to securely publish packages there
Mia@MiaAI_lab ·
64 GB DDR5 prices 💀
- elvis
@omarsar0 ·
Great paper from Amazon and colleagues on self-improving multi-agent systems.
Episodic memory for agents is so underrated.
But I am starting to see lots of good applications as frontier models improve.
This paper is the latest example.
They built a system that has three parts, all run by the same LLM. A controller plans each task and runs it one step at a time, choosing which sub-agent and which skills to use at each step. Sub-agents start as a generalist, and each specialist inherits the generalist's prompt plus its own role-specific method. Skills are short procedures, such as algorithms or pitfall checklists.
EMMA treats the controller's routing memory, the agent prompts, and the skills as episodic memory that improves from failures. After a failed task, a reflector decides which of the three owns the fix and proposes one edit there. Guards keep every accepted edit scoped, incremental, and bounded.
EMMA beats ACE, GEPA, EvoMAS, and ADAS on KramaBench, AppWorld, and BBEH, by up to 32.5% over the strongest baseline, and uses 17 to 38% fewer training executions.
https://arxiv.org/abs/2610.35966
Chat with Paper: https://academy.dair.ai/papers/emma-structured-episodic-memory-for-multi-agent-workflow-synthesis-2610.35966
Santiago@svpino ·
People are now writing code faster than they can roll it out to their users.
You need a plan for rolling out new features and updates. This is not optional anymore.
I've made so many mistakes deploying updates to software I've built that I could write an encyclopedia about everything you should and shouldn't do.
One of the most important lessons I've learned:
Never rush new code in front of all your users, especially now that agents generate so much of our code.
You need a way to decide what goes live, who gets it first, and when it's ready for more people.
This is how I usually implement this:
1. I have a new feature I want to release. I put it behind a switch in the code so I can turn it on or off. We call this "feature flags."
2. After I test the code, I deploy it with the flag off. Nobody has access to it at first. I want the deployment and the feature to become two separate decisions.
3. I then turn the flag on for a small group of users while everyone else keeps using the existing version.
4. I monitor what happens with the feature. Does it work as I expected? Are there any errors?
5. I gradually let more people use the new feature until I'm comfortable removing the flag.
6. If something breaks, I can turn the flag off for everyone and stop the problem until I fix it.
I call these gradual rollouts "canary releases". They're easy to implement with feature flags.
Feature flags don't replace code reviews or testing. They just give you another layer of control.
You don't have to build a feature flags system yourself. @LaunchDarkly manages the whole thing for you:
• Add flags to your code • Choose who gets the feature • Increase the group over time • Turn flags on or off without redeploying • Implement guarded rollouts to watch metrics during a release • Use automatic rollbacks
Here is a link to get started: https://fandf.co/4yWcjSq
Agentic coding is amazing, but it's only worth it if you can confidently put working software in front of your users.
Thanks to the LaunchDarkly team for partnering with me on this post.
Nathan Lambert@natolambert ·
Private benchmarks often show open models are further behind, but the problem with private benchmarks is you don’t have the community to squash the inevitable bugs for you. I'm 50/50 on if I think private or open-source is the best approach.
Aravind Srinivas@AravSrinivas ·
Fun project on Perplexity Computer: Asked it to design a GTA V mission where you vibecode GTA VI from an in-game phone while the devs call the cops on you for shipping GTA VI before them.
It went through the FiveM docs (natives, ped models, server convars, the limits of the in-game browser) and built the mission as a FiveM resource in Lua. Then it built an MCP bridge, a small local Node service that signs in to the Perplexity Computer MCP server with OAuth and turns your in-game prompt into a real Computer task.
The phone streams Computer's steps live during the chase, you answer its questions and approvals while driving, and you ship the build by losing the cops at the airport hangars.
For the voices, Computer rented a cloud GPU and ran Qwen3-TTS locally on it, with no TTS API, to voice every line of dialogue.
Derya Unutmaz, MD@DeryaTR_ ·
This is a very exciting potential new treatment for food allergies! But to appreciate just how remarkable this development is, let me share a personal story about Xolair, the anti-IgE antibody currently used to treat allergies, and how painfully slow biomedical progress has been until now. I believe AI is about to change this dramatically!
Back in 1989, 37 years ago, I was a visiting medical student as an intern at Ciba-Geigy immunology laboratories (which later became part of Novartis) in Basel, Switzerland. In the immunology department where I was working, one of the scientists, Christoph Heusser, had proposed developing antibodies against IgE, the molecule responsible for triggering allergic reactions, back in 1986!
Around the same time, scientists at a small biotech company called Tanox had independently developed a promising anti-IgE antibody. Ciba-Geigy and Tanox subsequently formed a partnership in 1990 to advance this approach.
I still remember the excitement surrounding this research as vividly as if it were yesterday! Imagining the possibility of treating allergies by targeting a single molecule responsible for triggering the entire allergic response was extremely novel back than. I was fortunate enough to witness the early development of this revolutionary idea firsthand, and it was incredibly exciting!
But here is where the story becomes rather astonishing.
Genentech subsequently developed a similar antibody in 1993, and the three companies eventually joined forces. Their collaboration led to Xolair (omalizumab), which finally received FDA approval in 2003, 17 years after the original concept! And that approval was initially for allergic asthma. It took another 21 years, until 2024, for Xolair to receive approval specifically for food allergies!
Think about that. Almost 40 years after those original discoveries, Xolair remains the only FDA-approved anti-IgE antibody for food allergies. That is how painfully slow the traditional process of developing new medicines has been!
Now compare that timeline with what Claude and Divergence Labs have accomplished in approximately ONE MONTH!
Claude generated roughly 35,000 antibody designs, helping identify promising candidates that could potentially remove IgE already attached to immune cells. This could be a major advance over Xolair, which primarily neutralizes free IgE. Some of these new antibody designs already performed better than Xolair in specific laboratory tests!
Of course, these are still early experimental results. The complete therapeutic antibody needs to be developed, validated, and eventually tested in clinical trials.
But consider the implications!
We are witnessing the beginnings of a fundamental transformation in biomedical research. As AI intelligence continues to advance exponentially, combined with automated laboratories, virtual cells, and eventually patient digital twins, I believe we will compress decades of drug discovery into months or even weeks!
Clinical trials and regulatory approval will also become much faster as AI improves our ability to predict efficacy, toxicity, and treatment responses.
My prediction is that instead of waiting decades for individual breakthrough medicines, we will eventually be developing hundreds, perhaps thousands, of new treatments in parallel, within just a few years!
Having personally witnessed the beginnings of anti-IgE antibody development nearly four decades ago, and now seeing AI begin to revolutionize that very same field, makes this moment extraordinarily exciting for me.
I have been waiting for this transformation for most of my scientific career. And I believe we are finally entering the era when the pace of medical discovery will accelerate beyond anything we have experienced in human history!
- Daniel San
@dani_avila7 ·
Firecrawl MCP is now available on http://aitmpl.com
You can use it with Claude Code to scrape websites, crawl entire sites, search the web, and extract content as Markdown or structured JSON
It includes tools like scrape, crawl, map, and search
Claude Code can call them directly when needed
Install it with one command:
npx claude-code-templates@latest --mcp devtools/firecrawl
You’ll need a Firecrawl API key
Maziyar PANAHI@MaziyarPanahi ·
I trained a small open model overnight to say "not sure, ask a doctor" 🔥
it answers on its own only when it's 90%+ sure, and hands everything else to a human
when it answered alone, its mistakes dropped from 46 to 13
which decisions would you let an AI make alone?
Vox@Voxyz_ai ·
If you still think AI video 𝗹𝗼𝗼𝗸𝘀 𝗳𝗮𝗸𝗲, look at the bottom half.
That's FLUX 3 Max, now live in Arcads. The realism is insane..😭 You'd swear it was real footage.
It's also up to 𝟮𝟮𝘅 𝗳𝗮𝘀𝘁𝗲𝗿 than other leading video models at the same quality. A 10-second clip comes back in about 12 seconds.
At that speed, you can try several camera moves on the same shot and keep the best one. Call the move by name in your prompt:
→ dolly zoom: the person stays the same size while the background stretches away or closes in → rack focus: focus shifts from the foreground to the background → FPV drone shot: the camera rushes forward like a drone → whip pan: the camera snaps sideways and blurs into the next shot
When you name one, add the seconds and the speed, like "dolly in, 3 seconds, slowly toward the person."
- Nathan Lambert
@natolambert ·
Open models' demise has been greatly exaggerated, but the previous AA index did flatter them. As an industry we need more specific intelligence indices, as LLMs are used in vastly different tasks.
My meta-review on the benchmarking industry. https://www.interconnects.ai/p/artificial-analysis-v5-and-the-benchmarkers
Marc Lou@marclou ·
And that, ladies and gentlemen, is the end of UI for SaaS
- Romain Huet
@romainhuet ·
Codex on the iPhone Duo is 🤩
Gergely Orosz@GergelyOrosz ·
I'm starting to sense a divide in the tech industry:
- People thriving on autonomy who are reinventing how they work and what tools they use with AI: LOTS of work, but they're enjoying it and getting very efficient
- People used to being mentored / trained / or simply told what to do and how. Frustrated with the lack of guidance from management on what to do with AI; also frustrated by seeing some of their peers thriving; increasingly bitter about AI and the state of the industry, as a whole
Neither group seems to be aware that there's another group, or have empathy for the other group