Rohit Kelapure

@claudeai is cooking hard today

how your set up your folder is the difference between you and the great karpathy. folders are everything. this is a great guide into building each layer of the wondrous folder. do bookmark this. and do implement this.

Top 10 Rules for reverse engineering COBOL code with LLMs 1. Invest 40% of Effort in Structural Parse Quality Before LLM Extraction 2. The Output Format Is the Specification — Define It Before Writing Prompts 3. Classify Programs Semantically Before Extraction — Not After 4. Ground-Truth Analysis Must Use Enriched Parser Output — Never Raw Source 5. Multi-Pass Extraction Must Have Bounded Feedback Loops with Circuit Breakers 6. Mandatory Code Review After Every Implementation Batch — Before Marking Complete 7. Copybook Resolution Is the #1 Technical Risk and Must Be Validated First 8. Every Number in Every Deliverable Must Trace to a Real Pipeline Run 9. Self-Correction Chains Transform LLM Extraction from Noisy to Reliable 10. Test-Driven Development Is Non-Negotiable — The Failing Test Comes First

The Municipal Police of Madrid has published the video of Paco's training exercises, the hero dog who saves and performs cardiac massage (CPR) on the officer who fainted (as a simulation), and it's winning hearts worldwide. ❤️

Really great writeup

Just setup autoresearch should be fun to play with ...

@karpathy you are making GPUs on VMs great again :-)

Just sent this to a buddy of mine youtube.com/watch?v=qb90PP…

@michalbrojak @trq212 Never

Five things they don't tell you about Forward deployed engineering in public ... (1) Forward deployed engineering means forward deployed i.e. at the customer site sitting down with the nurse and understanding if ambient listening summary worked. Working with the scientist or accountant who is using your replacement of their excel workflow for general ledgers or study management. It is not sitting in the armchair slinging features with cursor and claude. (2) Forward deployed engineering is successful because that operational leader who came up through the ranks with an associate degree and is now a VP is taking a blind all-in bet on your AI project with zero priors. Its the nurse or the audit specialist helping you automate workflows with their own job at risk. It is not because you were a cracked engineer or know how to write AI agents. (3) Forward deployed engineering is the painful work of grinding through a six month hill climbing eval step by step to improve the performance of your fax/image scanning LLM solution from 99% to 99.9% and then to 99.999%. Its not your magic PRD or LLM-as-judge eval that your wrote on the weekend that will get it done. (4) Forward deployed engineering ends in getting the job done, not 0 to 1. Its making the software useful and good. Its onboarding and reducing the friction for adoption. Getting it in the hands of practioners and driving business impact. Building is cheap, adoption and change management is HARD. (5) Forward deployed engineering is battling excel and ossified procesess. Its about dynamic, simpler and faster product that solves the user problem not time/money consulting or software-as-a-service. Its is not about solving the hardest problem - it is about solving the customers problem taking as long as it takes. AI was not used in writing this post. #FDE

Adversarial interoperability Lets go! claude.com/import-memory

> you work 9-5 every day > build your AI side project at night > “I will escape the underclass” > then realize you’ve just been working a second unpaid shift as LLM data generator, only this job sends you a $200/month invoice.

This Speech >>>> ❤️❤️❤️😭😭😭

Plan of plans developers.openai.com/cookbook/artic…

Evening vibed a system that provides an AI powered document analysis & rule-based screening for clinical diagnosis

Tenía 39 años y todavía tenía que revisar mi cuenta bancaria antes de invitar un café. No era un “artista joven y bohemio”. Era un hombre de mediana edad, cansado, sirviendo mesas en Nueva York mientras veía cómo mis amigos triunfaban, se casaban y compraban casas. Yo solo coleccionaba rechazos. Diecisiete años audicionando. Diecisiete años escuchando “no eres lo suficientemente latino para este papel” o “eres demasiado latino para este otro”. Hubo una noche en particular, después de que despidieran a mi agente, en la que me senté en el borde de mi cama y pensé: “Ya basta, Pedro. Esto no va a suceder”. Sentí que le había fallado a mi madre, que se había ido demasiado pronto y nunca me vio llegar a ninguna parte. Hoy, mi cara está en una pantalla gigante en el medio del Super Bowl, anunciando el final de temporada de la serie más vista del mundo. La gente me llama “el novio de internet” y los directores que antes ni me recibían, ahora me mandan guiones sin que yo los pida. No escribo esto para presumir mi éxito. Lo escribo para decirte que no existe tal cosa como “llegar tarde”. Si hubiera triunfado a los 20, la fama me habría destruido. Llegué a la cima cuando por fin aprendí a escalar sin mirar hacia abajo. Si sientes que se te está pasando el tren, recuerda mi cara de cansancio a los 39 años. Tu historia no ha terminado, apenas está en el segundo acto. Pedro Pascal

Here is my @claudeai personality. What is yours ? @AnthropicAI

I am training @openclaw bot as a mirror image of me. In case I die, Rushil has a backup.

ok I a in ... ✓ Detected: macos ✓ Homebrew already installed ✓ Node.js v22.9.0 found ✓ Git already installed → Installing OpenClaw 2026.1.30...