Jeevesh Jain

Composio is now available as a Cursor plugin. Connect your agent to 1000+ apps, securely. ⬇️ try it out

Link to apply - jobs.ashbyhq.com/composio/cceb9…

Hiring a Product Intern @composio. College no bar, 2 or 6 months depending on availability and candidature. The expectation is being a cracked builder, get things done with agents and be in our Bangalore office. You will be working closely with me and @kalapolish Link to apply below 👇

New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.

By the way, I used Wispr to create this, so I'm not saying the product is bad. It's actually really nice, one of the few products I've liked in the past 1–1.5 years, because Claude Code is taking most of my other tools away.

I'm trying to adopt @WisprFlow for content creation, but it feels subpar. The transforms feature is still in beta so my guess is it will improve, but right now I think I could get better results by typing. I’d like to hear from the Wispr product managers about the best way to use the tool. It seems no deep research is performed when I write content; the system simply takes my input as‑is. Could we add an option to enhance output with web search or research instead of only transforms? Can we use a skill instead of a simple prompt?. I’d appreciate clarification.

We tested GLM 5.2 against Claude Opus 4.8 and GPT-5.5 on 41 agentic tasks that use real tools like GitHub, Jira, and LaunchDarkly. GLM tied or won on every task. On one, it was the only model to get the task right. The task was to find stale feature flags in LaunchDarkly, a tool for managing feature flags. A flag counts as stale only if it's switched off and nobody's planning to touch it. There were two flags, and both were off, so at a glance both looked stale. ..except that one of them wasn't . That flag had a change queued up and waiting for approval to switch it back on, which meant that technically it wasn't stale. The only way to identify this was to check the pending approvals list, not just read the on/off state. GLM 5.2 made that check and got the task right. Opus 4.8 and GPT-5.5 both stopped at on/off. Final scores across all 41 tasks: GLM 5.2 got 40 out of 41. Opus 4.8 and GPT-5.5 each got 39.

Saw this update on Claude Code, I have done CLI and terminal PMing before, it is so hard to decide what can be built. I remember hearing @bcherny talk about this in his YC podcast some time back. Terminal design is one of the hardest UX changes, For those of us who live in the terminal instead of a desktop app, small efforts like this matter a lot.

new chapter. building agent infra at @composiohq as a founding PM (Builder). the boring stuff nobody wants to build is usually the stuff everything runs on. that's the bet. excited to work with the 🐐s @KaranVaidya6 @GanatraSoham @kalapolish 🚀

> be me > maxed out on Claude Max 20x > think my wallet has suffered enough > @AnthropicAI ships Fable 5 > best model they've ever made public > free on Max til June 22 > then "usage credits required" > start typing angry comment > read their reasoning first > "demand is unpredictable, we'd rather ship now than make you wait" > "we'll put it back in plans once we can handle the load" > delete angry comment > came to rage, left nodding still gonna pay though. we always pay

@karpathy joining Anthropic is the AI version of Raghav Chaddha joining the BJP.

last week in tech was based.

Kimi K2.6 now ranks 4th on Artificial Analysis Intelligence Index and here is why that is crazy 👇 - Costs 5-6x less than Opus 4.6 (which is a strong model) - Beats Opus 4.6 in 6 categories, including the agentic index, and stays close on almost all the others - Smaller model, capped at 256k context vs 1M for Opus 4.6 - Slower than Opus 4.6 - But the craziest part, it is an open weight model. We are now in an era where open weight models sit less than 2 months behind the closed frontier.That is good news for everyone. Inference costs drop. Self hosting becomes viable again. Agent workloads stop being coupled to one provider. And maybe, just maybe, this is the start of a new pseudo on-prem era for agents.

Top 4 worst hiring experiences so far in my career. - You want us to pay you this much AND train you? (I had said I wanted to learn from the leaders in a DeepTech sector.) Got the offer. Rejected it. - Got asked ChatGPT'd questions with no substance, by someone I could tell did not understand what they were asking. I almost laughed. - Found out I was competing with two Staff PMs with 15+ years of experience. Same industry. (This happens a lot. Almost a lot of rejections I've had is because someone multi levels senior was in the final rounds. Decently motivating.) - Interviewer did not show up after rescheduling thrice. (This has surprisingly happened to a lot of people I know. Not really the best behaviour.) What are yours?

YC Startup School India edition: Three founders, two IITB CSE folks, and one VC walk into a samosa line. Nobody gets samosas. Peak Bangalore. #yc #bangalore

Me after opus 4.7

Here's how my friend made $500k in 1 hour of algorithmic trading using Claude Mythos Preview. He's among the select few with access. Here's the exact setup: - Crafted a custom skill for Claude Code, specialized in algorithmic trading. Not just any skill. A specialized one. - Built a second brain using Karpathy's LLM wiki idea. Interconnected knowledge graph. Every alpha-generating technique, wired together. - Asked Claude to run simulations. Then simulations of simulations. Then picked the best one and deployed it to real markets. He gave the system $2 million. It has $500k left. The man is a genius. The system worked flawlessly. Claude Mythos Preview is clearly the future of finance.

@jenzhuscott @karpathy I think the next big step essentially is the LLM driven software to hardware intersection getting democratised. Agents want to use the computer now, I wonder how far are we of making them use other hardwares and essentially giving them real world hands (and brain)

As I build my own 2nd brain 🧠 on Obsidian using @karpathy ‘s wiki idea, it suddenly dawned on me - one day when we r gone, our kids could inherit an interactive map to your mind, passion, obsessions, work, fascinations… It’s kind of beautiful way to think abt your 2nd 🧠.

@karpathy Folks in non-tech jobs obviously, but also engineers in MNCs and service companies with restricted access are still on archaic setups and still calling it overhyped.

Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code. But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along. So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions. TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.