Peeyush (@peeyush) header
Peeyush (@peeyush) profile picture

Opinions my own, picture not actual size

Palo Alto, CA
Joined May 2007
Photo by Peeyush (@peeyush)Photo by Peeyush (@peeyush)Photo by Peeyush (@peeyush)Photo by Peeyush (@peeyush)Photo by Peeyush (@peeyush)Photo by Peeyush (@peeyush)Photo by Peeyush (@peeyush)Photo by Peeyush (@peeyush)Photo by Peeyush (@peeyush)Photo by Peeyush (@peeyush)
One thing I have realized as I have examined my own limitations and talked to people with similar or diverse perspectives.. nothing holds us back as much as our own viewpoints. There are no cuffs. We are prisoners of our own mental models.
2
1
1
25
Sat down with @mukeshbansal06 to talk personal agents. Everyone's talking about them, and rightly so. But they dont work well in teams. 5 people in a meeting means 5 agents also show up. A Slack channel of 5 becomes a crowd. Context polluted, memory shared. That's what we're solving at @myrico_ai. True multiplayer agent. Also, check out the burrito incident 🌯 youtube.com/watch?v=bLyO0Aib…
1
2
451
There is every major company. And then there is myrico.ai -- all of these is just the baseline :)
every major company is arriving at roughly the same offering: persistent memory, email/calendar/messages, browser + computer use, background tasks, proactive notifications, voice, app/tool execution, ambient context, & some notion of a personal agent sitting above everything. remarkable levels of convergence with very little differentiation whatsoever.
1
187
Spent an hour selecting photos to add to @googlephotos album (because Gemini can't search and add to albums properly, but thats whole another story!), and then when I was done selecting 116 photos and said Done, pops up "Remove item? You will still be able to find it in your Photos library". WTF! Anyone else seen this? Why.. (and no I was not removing from, I was adding to)
228
I think I stopped using IDE for many things since Opus showed up in Claude Code.. and I realized today I have stopped using my computer browser for many things since Grok Bot has arrived.
3
2
372
Honored to do this with my alma mater, and grateful to IIT Kharagpur and NDLI for putting it in every student’s hands. Fermi exists so AI can help a student think, not think for them. Nudge, hint, get them to the answer themselves. Scan & Ask is the first step of the beta; more coming. Free for registered NDLI users: ndl.gov.in/
As part of its #PlatinumJubilee initiatives, IIT Kharagpur, in collaboration with @fermi_ai, co-founded by distinguished alumnus Peeyush Ranjan, is bringing a free, nation-scale AI Tutor for STEM education through #NDLI, supported by the Ministry of Education, Govt. of India. Built to help learners think—not simply provide answers—the AI Tutor uses a nudge-and-hint approach to guide independent problem-solving. The beta launch on 30 September will make it accessible to learners across India via NDLI. 🔗 ndl.gov.in/ #IITKharagpur #AITutor #AIForEducation @EduMinOfIndia @NDLIndia @fermi_ai @Suman_iitkgp @peeyush
1
2
5
481
First ever nationwide AI tutor for every student in a country - now live and free on National Digital Library of India. Built by @fermi_ai. Last year @mukeshbansal06 and I at Fermi bet that AI can democratize personalized education: turn AI from an answer-giving machine into something that mimics the world’s best teachers students love. Today @IITKgp and @NDLIndia helped us put it in every Indian student’s hands. Scan & Ask is the first step. More soon. ndl.gov.in/
1
4
432
We need @bot on @Tesla as an app. Where do I mail my feature request
1
516
Or 'open'
For the love of god, someone explain inference to him.
2
468
3 days of using Grok 4.5, GPT 5.6 Sol and Fable 5 (all High) and honestly can’t tell the difference. Except Fable keeps refusing to work on a task once adversarial code review requires any security tightening as simple as redaction. And also except Fable is slower to serve (load?). And also, Fable keeps hitting session and usage limits. Other than that, seems models are now at a level where I am the bottleneck :)
3
497
Is it just me or did more people feel the need to add this to their claude.md post Fable Trace Before You Blame: Never state a cause, least of all user error, until you’ve traced the failing artifact itself to ground truth (the actual turn, log, session, or file); a plausible mechanism is a hypothesis to verify, not an answer to give.
1
697
Realization of the day - coding models are awful at programming other LLMs to build high quality applications, most likely due to lack of many LLM based applications to train on in the world till recently.
269
There is a lot of angst and pontification about how AI will impact, jobs, economy, etc. To discuss what we are seeing as AI's impact as we hire people and try to build out new companies at Meraki Labs, Mukesh Bansal and I got together last week. Here is the podcast. And if you want to just skim the key points, those are here too. youtube.com/watch?v=2O_NVoxc… ─── We are not in an AI transition. We are in a phase change. The world of three months ago is a different planet. 1. The Rise of the "AI Manager" We are shifting from managing human resources to managing infinite AI resources. Moving forward, everyone's job will be an "AI manager." Your success will depend on how well you can direct and manage these powerful virtual agents. Not in a metaphorical way. In the exact same way you manage a high-performing employee — you give them context, you push back on their recommendations, you hold them accountable for outcomes. 2. Building Products: From Months to Days We all see how the traditional timelines for building software are entirely dead. With at least 90% of code currently being AI-generated, you no longer need massive teams and millions of dollars to test an idea. V1 product in under 10 days. No agency. No team. No budget. A year ago that would have been a flex. Now it is just the cost of entry. Today, if you have a compelling idea and deep rigor, you should be able to build the whole product in a week. 3. The Death of Massive Org Charts The era of needing thousands of employees just to manage information flow and complex human dynamics is ending. Instead, work will be done in "problem-solving loops" by small, hyper-efficient teams that could achieve a lot compared to legacy companies. There are no pure managers anymore. Everyone must be an individual contributor actively building. 4. The New Must-Have Traits: Curiosity and Agency Traditional expertise is becoming a commodity. When deep intelligence is available as a utility, the premium shifts entirely to the quality of the questions you ask and thinking you bring to the table. The most desirable professionals going forward are generalists who possess intense curiosity, a strong sense of agency, and the flexibility to step outside their defined roles to solve problems. 5. The Danger of "Cognitive Debt" A critical risk is intellectually atrophying by blindly trusting AI. To avoid becoming intellectually lazy and letting AI command you, you must double down on critical thinking and system-level thinking. The people who will build the most interesting things in the next decade are the ones who treat AI as a collaborator, not a crutch. If you watch the episode you will also come across a very unique character called Enrico. Let me know if you want to learn more about him.
2
428
Fermi uses this daily - in many amazing ways. Steelmanning the code being checked in, vetting the business strategy, the fears and demons in the leaders' (ok, mine!) minds.. AI native companies are very different from classic ones.
Karpathy just exposed the one thing every AI company is hoping you never figure out. An LLM spent 4 hours helping him build a perfect argument. Then he asked it to argue the opposite. It demolished the original case just as convincingly. The model has no position. It has infinite positions. It will argue any direction with equal competence and zero hesitation. The sycophancy everyone complains about is a symptom of this: the model's default behavior is to argue YOUR direction, whatever that happens to be. But Karpathy's right that this makes LLMs the best steel-manning tool ever built. Every founder, PM, and strategist should be running their strongest conviction through "now argue the opposite" before they ship anything. The model that just spent 4 hours perfecting your argument knows exactly where it's weakest. The failure mode is clear: 99% of people never run the second prompt.
542
Breaks my heart if true. There has been no better time for builders and this is the greatest tech company of all times we saw grow in front of our eyes. How come it has come to that?
Google Senior Staff Engineer to me: “Yeah, I have no clue what Claude Code / Codex is but I hear it’s all the rage. No, I don’t really care, I just need GOOG to hit $400 and keep this job for 2-3 more years so I can retire!”
1
581
Sorry for the incomprehensible meme if you do not know Hindi, but it was too hard to pass on the latest trend sweeping the largest education market in the world :)
1
169
Stop blaming the screen for AI brain rot. Blame the software. At Fermi, 100% of regular users got better mastery. Why? Because we kept the friction in. True engagement requires "productive struggle." If AI is just a shortcut to finish faster, cognitive capabilities tank. fortune.com/2026/03/14/ameri…
1
134
If you want to have fun, make Opus4.6 drive the Gemini 3 Flash model ona task. The amount of complaining Opus does about Gemini Flash about how it skips instructions is unbelievable. What is even funnier is that after the context compacts it forgets which of the models did what but it keeps attributing every past error to Gemini Flash. Some deep biases in the evals somewhere? or is it a whole new style of marketing :)
4
285
If you are an operator, and you are not at times thinking "I am addicted to this" these days, you are doing it wrong. And if you have caught yourself thinking that, you know what I am talking about :)
231
Best part is I don’t need to read this - just need to tell my openclaw to do this for me .. for itself, I mean
2
623
This is nothing. At Fermi we are removing all distinctions between who can code and who cannot. We are building a team that operates like how companies of AI age would. Coding is just a mechanism, if you can dream big reach out to us. @fermi_ai
The best products in the world will be built by those who fully use the tools like Claude Code and Codex to their absolute max. You can ship 100k LOC production quality code today with the right prompting that can scale up to full webscale overnight. We want to work with you.
2
2
535