San Francisco, CA
Joined January 2018
anish retweeted
Today, we’re releasing Continual Learning Bench 1.0: the first, realistic benchmark for measuring how AI systems can improve in online settings. Benchmarks today assume models are stateless. Each example is independent, and once a system finishes a task, it moves on as if nothing happened. But deployed AI systems should learn from experience. We tested 10+ frontier systems against novel, expert-validated tasks and find there’s still plenty of headroom for learning. (1/n)
42
165
47
1,188
845,047
anish retweeted
Introducing Flue — The First Agent Harness Framework Flue is a TypeScript framework for building the next generation of agents, designed around a built-in agent harness. Flue is like Claude Code, but 100% headless and programmable. There's no baked in assumption like requiring a human operator to function. No TUI. No GUI. Just TypeScript. But using Flue feels like using Claude Code. The agents you build act autonomously to solve problems and complete tasks. They require very little code to run. Most of the "logic" lives in Markdown: skills and context and AGENTS.md. Flue is like Astro or Next.js for agents (not surprising, given my background 🙃). It's not another AI SDK. It's a proper runtime-agnostic framework. Write once, build, and deploy your agents anywhere (Node.js, Cloudflare, GitHub Actions, GitLab CI/CD, etc). We originally built Flue to power AI workflows inside of the Astro GitHub repo. But then @_bgiori got his hands on it, and we realized that every agent needs a framework like Flue, not just us. Check it out! It's early, but I'm curious to hear what people think. Are agents ready for their library -> framework moment?
191
340
105
3,817
800,728
For the agentic economy to overtake the human economy, agents need a way to discover services. We launched Agentic(.)market to give agents a discovery layer to find and integrate x402 services seamlessly. Add the skill to your agent. And list your services to start earning revenue. > npx skills add coinbase/agentic-wallet-skills
Introducing Agentic(dot)Market, the homepage of the agent economy. - Monitor agentic commerce trends - Discover services for your agent to buy - Sell your services to agents Thousands of services. Zero API keys. Powered by x402.
266
240
72
2,253
510,364
Introducing: Browser Harness. A self-healing harness that can complete virtually any browser task. ♞ We got tired of browser frameworks restricting the LLM. So we removed the framework. > Self-healing — edits helpers. py on the fly > Direct CDP — one websocket to Chrome > No framework, no rails, complete freedom > Drop-in for Claude Code and Codex I challenge anyone to find a task that DOESN'T work. I couldn't yet.🔥 100% open source ↓
175
291
63
3,566
1,088,490
We created the open source version of Claude Managed Agents. Introducing Multica github.com/multica-ai/multic…
Introducing Claude Managed Agents: everything you need to build and deploy agents at scale. It pairs an agent harness tuned for performance with production infrastructure, so you can go from prototype to launch in days. Now in public beta on the Claude Platform.
91
198
57
2,191
446,676
Excited to announce Juicebox has raised $36M in funding, including our Series A led by Sequoia Capital (@sequoia) - to help the world’s leading teams win the talent war. More below 🧵
50
43
29
488
206,928
Announcing Pylon's Series B led by a16z and BCV Since starting the company 2.6 years ago, 780+ of the fastest growing companies including Together AI, Cognition, and Temporal have chosen to scale their customer support on Pylon.
58
24
15
563
129,109
Introducing Account Intelligence Turn customer data → signals → action. The result is fewer blind spots, faster reactions, and proactive workflows that drive retention and revenue.
13
11
4
79
16,679
Join a startup before starting your own... Here's why 👇 1/ Learn faster You'll expedite your learning by working with people who are a few years ahead of you. Specifically around their pace and decision making process. You'll see everything from how they launch products, hiring, think about markets, culture, build product, etc. 2/ It will build your hiring pipeline It's incredibly hard to hire people (especially if you're a younger founder). No one gets excited about working for someone younger and less experienced than them. Instead, if you join an early-stage company you'll meet other great people who you can try to recruit after you leave. 3/ Investors Early-stage investors will care a lot about where you've worked and social proof. Having your former founders tell an investor you're good will help a LOT in the long-term. -- Notes -- - Important: The type of company to join is one between 30-100 people, Series A-C, less than 3 years old, backed by Tier 1 investors (a16z, Sequoia, etc), has grown revenue minimum 3x in the past 12 months, minimum of $3M+ in ARR (you want them to have early PMF), under $1B valuation, and importantly works together 5 days a week in person. - You'll be bad at picking which startups are good to join. This is one place where good investors can help you out. Pick a good VC and email a partner there saying something like "I'm graduating, want to grind, and start a company long-term. Are there any portfolio companies you'd recommend I take a look at?" They will be a decent filter and can also give you their take on other companies you might be looking at. - If you do have multiple offers and are unsure who to join just pick the company that seems to have smarter people. I'd optimize for intelligence of the team over things like current revenue scale. -- If all of this excites you and you want to start a company one day, checkout Pylon. We're growing incredibly fast and hiring across all roles... - Sales - Marketing - Support - Customer Success - Engineering - Design - Ops - Recruiting Jobs link in the comments.
2
1
1
26
1,593
Today, we're announcing our $32M Series A at a $300M valuation led by Insight Partners. Delve is automating busywork for humanity, starting with compliance. Here's our story.
156
57
57
1,228
459,833
We're excited to finally share what we've been up to :) Announcing Endex: An AI Financial Analyst Today we're coming out of stealth and announcing our partnership with @OpenAI 🧵
91
84
28
1,619
315,433
I’m extremely proud to share that @glean surpassed $100 million in ARR in our last fiscal year in just over three years. I want to start by thanking everyone who has supported us along this journey – our customers, partners, awesome end-users, supporters, and team. I am so grateful for your support.
33
26
9
728
89,664
Accepted to ICLR 2025!
Looking into the inner workings of VLMs can really change the game in improving model reliability. Thrilled to see this work with @nickhjiang and @YGandelsman out in the world—check out our latest paper!
1
16
1,310
🔍 Curious about what's really happening inside vision models? Join us at the First Workshop on Mechanistic Interpretability for Vision (MIV) at @CVPR! 📢 Website: sites.google.com/view/miv-cv… Meet our amazing invited speakers! #CVPR2025 #MIV25 #MechInterp #ComputerVision
14
6
57
24,765
Huge. UC Berkeley just released a $450 open-source reasoning model that matches o1. Sky-T1-32B-Preview is a fully open-source model designed for reasoning and coding tasks. Achieves 82.4% on Math500 and 86.3% on LiveCodeBench-Easy. It includes training data, code, and model weights. Open-source resources: > Training infrastructure and scripts. > Full technical report with wandb logs. > Model weights for Sky-T1-32B-Preview.
61
201
31
1,835
170,965
Have you ever wondered why we don’t use multiple visual encoders for VideoLLMs? We thought the same! Excited to announce our latest work MERV, on using Multiple Encoders for Representing Videos in VideoLLMs, outperforming prior works with the same data. 🧵
5
15
3
128
18,265
Update on @suchirbalaji We hired private investigator and did second autopsy to throw light on cause of death. Private autopsy doesn’t confirm cause of death stated by police. Suchir’s apartment was ransacked , sign of struggle in the bathroom and looks like some one hit him in bathroom based on blood spots. It’s a cold blooded mu*d*r declared by authorities as suicide. Lobbying in SF city doesn’t stop us from getting justices. We demand FBi investigation @elonmusk @VivekGRamaswamy @GavinNewsom @MEA
567
4,273
466
21,459
2,894,376
The "Vision Transformers Need Registers" paper by @Meta is now in @huggingface Transformers🤗 Similar to "thinking tokens" in LLMs, vision transformers that leverage "register tokens" perform better, and enable interpretable attention maps Links in second tweet!
5
69
2
516
36,346
Scaling interpretability work for frontier AI alignment has never been more important. That’s why we’re launching Ember - the first hosted mechanistic interpretability API with fast inference support for generative models like Llama 3.3 70B.
14
85
30
550
157,037
META JUST KILLED TOKENIZATION !!! A few hours ago they released "Byte Latent Transformer". A tokenizer free architecture that dynamically encodes Bytes into Patches and achieves better inference efficiency and robustness! (I was just talking about how we need dynamic tokenization that is learned during training 🥲 It's like fucking christmas!) I don't want to talk too much about the architecture. But here's a nice visualization from their paper. Let's look at benchmarks instead :) "BLT models can match the performance of tokenization-based models like Llama 3 at scales up to 8B and 4T bytes, and can trade minor losses in evaluation metrics for up to 50% reductions in inference flops!" This is basically a perplexity vs training flops chart - scaling laws with compute. BPB is a tokenizer independent version of perplexity. BLT is on par or better than LLama 3 BPE! Most importantly they scale this approach to train Llama-3 8B model on 1T tokens which beats the standard Llama-3 architecture with BPE tokenizer!
92
680
138
5,110
490,472