Manasvi Kapoor

we cracked it. the cch= signing system in claude code is fully reverse engineered - all credits for the work go to @ssslomp who did an amazing re work now any opensource client can let users actually use the anthropic subscription they already paid for. with whatever tool they want already merged into free-code. third party clients can generate valid cch= hashes now, no official binary needed the "native attestation layer" lasted about a day lol

Just migrated our stack from vercel to cloudflare Feels like a burden off the shoulders

@pariscestchiant itshelm.com

i need someone to lend me $5,000,000 so i can buy helm(.)com our current domain is helmkit . com and im tired of people thinking its called helmkit it’s called helm i critically need a better domain name

@figma has so much potential to be such a great software both for designers and developers But for now, its just infuriating smtimes to use as a dev

Amaze amaze amaze

@pariscestchiant How's the progress going

spent 3 hours making a progress icon today ama

@aurorascharff Nicely timed

@void_exception Yes! Stay tuned!

What are you struggling with in Next.js right now? I work on Next.js developer experience at Vercel. Want to know what actually trips people up so I can help fix it and create content around it. Drop your pain points.

@chatgpt21 It's arguably that instead of getting spoilt there's a high chance the models themselves will get poisoned overtime inevitably with the test data

WOW! Models preform HORRIBLY on ARC AGI 3 Gemini 3.1 pro 0.37% GPT 5.4 (High) 0.26% Opus 4.5 (Max) 0.25% I wonder how long It’ll take for this benchmark to be solved

LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below

@pariscestchiant everyone is

so i’m just a crm entry to you after all huh?

Okay assume is this way A model can understand what a word, a sentence or a paragraph means, but contextually It never actually see's the exact letters That's what tokenization does It converts words, sentences and even punctuations with context to each other as set of numbers proceedings.neurips.cc/paper_files/pa… Highly recommend this for a bit more information about how tokenization and attention mechanism works in LLMs It's a great read and will tell you why llms are inherently flawed for questions like "How many rs in strawberry", they never see the word strawberry as a word with letter but would probably see it as a sequence of numbers like 2948489393934947 with no correlation to the letter "r"

@void_exception @opencode I was thinking the problem would have been fixed with this bullet point - Self-evolution - do a task > check results > fix mistakes > try again

MiniMax M2.7 available in Go - Better at complex tasks over M2.5 - Fast - give it a plan and it runs with it - Self-evolution - do a task > check results > fix mistakes > try again

@nearbycoder @opencode Because they're fundamental to how tokenizers work and largely are useless to evaluate a model

@opencode why are we back to these trivial problems with models?

We need like, --kinda-dangerously-skip-permissions I dint give a fuck about whatever long tail esoteric git command you've cooked up for some reason, just don't rm -rf my home dir

@iyoushetwt Screenshots

Frontend has screenshots. Backend has… what? How do backend developers show proof of work?

Doesn't even take open-source community 24 hours to release alternatives github.com/vxcontrol/pent…

I’m leaving the AI Impact Summit with fewer questions about models and more questions about system ownership. The most interesting conversations this week weren’t about capability. They were about responsibility. Who owns the output of an AI system? Who audits it? Who is accountable when automation becomes decision-making? As AI moves from assistance to autonomy, the technical challenge changes. It’s no longer about making systems smarter. It’s about making them explainable, observable, and governable. That shift is subtle, but it’s massive. Because intelligence without accountability does not scale inside enterprises. What stood out most this week was the seriousness of teams thinking about this early. Not how to ship faster. But how to ship responsibly. The summit may be ending. But the real engineering challenge is just beginning. Building systems that can be trusted, not just admired. #ResponsibleAI #EngineeringLeadership #AIInfrastructure #FuturixAI

The final two days of the AI Impact Summit were less about presentations and more about practical questions. The conversations shifted toward real deployment challenges. Documentation bottlenecks. Legacy systems. Integration into existing workflows without breaking stability. One thing became clear. AI adoption does not stall because models lack intelligence. It stalls because infrastructure struggles with reality. Across discussions with delegates, the recurring theme was operational friction. Static documents that slow down processes. Manual verification loops. Data locked inside PDFs and scanned records. Demonstrating intelligent OCR powered by Shivaay was less about showcasing capability and more about solving a concrete problem. Turning documents into structured, usable data inside real workflows. What stood out over these two days is that production AI is no longer experimental. It has to coexist with legacy environments, compliance requirements, and imperfect data. The next phase of AI will not reward the smartest model. It will reward systems that handle edge cases, drift, and scale without losing reliability. The summit conversations are ending. The architectural work is just beginning. #EngineeringLeadership #AIImpactSummit

San-script 🖌️