🚀 We're #hiring a Research Scientist – RL Training @SnorkelAI.
We need someone who's actually RLFT'd agents using complex environments (e.g. SWE-Bench/Terminal-Bench).
Deep hands-on experience with GRPO, RLHF, DPO, reward modeling & frameworks like verl/SkyRL. 30B+ scale and deep expertise in RL algorithms.
Come build SOTA coding agents with us!
📍 RWC / SF / NYC / Remote - US
#ML #ReinforcementLearning #PostTraining
Whether or not Europe stands with us, whether or not your journalists do their jobs, whether or not your politicians demonstrate the courage to act, I will fight for my people and my country.
One thing I’ve started doing recently (and highly recommend): Always ask your coding agent to create a security plan and maintain it in a local security.md file.
With recent security incidents affecting tools and platforms like Vercel and libraries like LiteLLM, it’s a good reminder that even modern stacks aren’t immune.
I spend ~30 minutes every now and then and ask my coding agent to review security.md and update it based on recent code changes and new risks
It’s a small habit, but it forces you to stay intentional about security as your codebase evolves.
If your app ever reaches prod, this can easily save you from expensive mistakes.
BTW, Iranian people (the cool ones, not the assholes), good luck, really. Things going well for you is things going well for everyone. I hope I'll discuss AI and startups and cool shit in a few years from now in a trendy coffee shop downtown Tehran.
Ukrainian President Zelensky:
There was so much talk about the protests in Iran, but they drowned in blood.
The world has not helped enough the Iranian people; it has stood aside.
By the time politicians started forming positions, the Ayatollah had already killed thousands.
What will Iran become after this bloodshed? If the regime survives, it sends a clear signal to every bully: kill enough people and you stay in power.
Tens of thousands of people marched in the streets of Toronto in solidarity with the people of Iran!They are all chanting for regime change in Iran and supporting @PahlaviReza as their representative for the transition period! #IranRevolution2026#IranMassacre@CBCNews@CTVNews
I think it’s all about evals! The bottleneck of agentic AI is evals and getting access to a diverse dataset of un/successful sessions of agents solving real world problems combined with real human feedback is gonna be a game changer!
Manus great ARR is a proxy that shows they have lots of successful episodes of agents completing tasks and the diverse set of tools they have added to it adds to the diversity of tasks manus can complete!
The key here is diversity of tasks manus agent can do! Any big tech company with access to this data and LLM infra can use this dataset to improve their models for general agentic tasks!
LoRA makes fine-tuning more accessible, but it's unclear how it compares to full fine-tuning. We find that the performance often matches closely---more often than you might expect. In our latest Connectionism post, we share our experimental results and recommendations for LoRA.
https://t.co/fYV4FPi71m
Approximately 44% of U.S. unicorn startups (valued at $1 billion or more) were founded by immigrants.
As of 2024, 46% of Fortune 500 companies were founded by immigrants or their children, collectively generating over $8.6 trillion in revenue and employing millions.
Immigrants have founded 60% of the top U.S.-based AI companies.
🚀 JUST LAUNCHED: X Space Converter
Convert unlimited X Spaces into content:
- 99% accurate transcripts
- AI summaries
- Ready-to-post tweets
All with ONE payment for LIFETIME access
no subscriptions, no limits.
Grab yours here: https://t.co/CJ2AwVuK5T
A bit of @Flowjincom's journey so far 🧵
We shipped our first version in late 2022, starting with a simple MVP.
Today? 800+ paying users who transform their long-form content into high-performing short clips.
For the past year, our focus has been laser-targeted:
→ Improving our clipping logic to find the best moments
→ Gathering data on what clips actually perform
→ Building a no-learning-curve video editor with AI features
We intentionally avoided the Adobe Premier route.
Our north star: powerful enough for marketers, simple enough for anyone to use instantly.
The result? 800 paying users saving time while creating better content.
Now we're accelerating growth by doubling down on what worked: purpose-built tools.
We're breaking our product into specialized solutions:
• Podcast clip maker
• Reel generator
• Webinar-to-shorts converter
Each solving a specific content problem.
First stop: revamping our Twitter Space downloader.
This simple tool built in 2022 has quietly brought us 3K visitors and 30 new paying users monthly—without any optimization.
We're not just enhancing it—we're expanding the entire value chain:
Download → Transcript → AI Summary → Ready-to-post tweets
Taking a Twitter Space and transforming it into weeks of high-quality, shareable content.
In minutes instead of hours.
Why focus here?
Because everyone hosting Spaces faces the same problem: hours of valuable insights trapped in audio that most people will never hear.
We're solving that by unlocking those insights and making them infinitely more shareable.
I'll be sharing more about why we're building this and opening our waitlist soon.
We believe in creating tools that give creators their time back while amplifying their reach.
If that sounds interesting, stay tuned... 👀
Career update:
Super excited to join @Pinterest 📌 as a Senior Machine Learning Engineer!
Looking forward to working on some really cool and challenging ML projects!
New Anthropic research paper: Scaling Monosemanticity.
The first ever detailed look inside a leading large language model.
Read the blog post here: https://t.co/6RYwxt6nWI