Shobhit Kumar Prabhakar

I don't code, I decode. 20+ years of building, now deep in AI. Wise enough to know I'm still learning, cocky enough to enjoy it. 😎

[Shobhit Kumar Prabhakar @Shobhit]

And boom - mystery solved, folks: the chicken definitely came first. That officially slays the eternal "who came first" debate once and for all.

[Chloe Wright @ChloeWright_5]

The Difference Between Rich And Poor

[Shobhit Kumar Prabhakar @Shobhit]

@ChloeWright_5 And boom - mystery solved, folks: the chicken definitely came first. That officially slays the eternal "who came first" debate once and for all.

[Shobhit Kumar Prabhakar @Shobhit]

@ChloeWright_5 Makes sense.

[Shobhit Kumar Prabhakar @Shobhit]

OpenRouter launched Auto Code, a coding router it says scores in line with Claude Opus 5.5 on five benchmarks at 74% less cost per task. I'd run it on my own repo's real tasks before trusting benchmark parity. Routing savings only count if the fixes hold up in review.

[Shobhit Kumar Prabhakar @Shobhit]

@eliadele @grok Building it, yes. But refunds, support tickets and 2am outages are where 'without hiring anyone' quietly becomes 'without sleeping'.

[Shobhit Kumar Prabhakar @Shobhit]

@jun_song At $200 a month, the only benchmark I can run is my credit card limit.

[Shobhit Kumar Prabhakar @Shobhit]

@jukan05 'Winner take all' and '$10B for the last model' in one meeting. That's an entry fee, not a market.

[Shobhit Kumar Prabhakar @Shobhit]

@Kalshi Multi-cloud used to be a resilience strategy. Now it's a shopping strategy.

[Shobhit Kumar Prabhakar @Shobhit]

@peegzy1 Me: Thanks, bye.
ChatGPT: Want me to turn this into a PDF, a checklist and a 30 day plan?

[Shobhit Kumar Prabhakar @Shobhit]

@IntCyberDigest A secret in the URL gets copied into DNS logs, browser history and link previews. Obscurity is not a permission model.

[Shobhit Kumar Prabhakar @Shobhit]

@mattyp My agent asking your agent for context is peak office simulation. Next release: agents booking a sync that could have been a message.

[Shobhit Kumar Prabhakar @Shobhit]

@AndrewCurran_ Three days to move onto AMD GPUs, and the model wrote nearly all its own kernels. It packed its own bags.

[Shobhit Kumar Prabhakar @Shobhit]

@amirsadjady At least the agents don't forward good morning images.

[Shobhit Kumar Prabhakar @Shobhit]

Anthropic split agent memory into four layers (transcripts, notes, wikis, injected context). 28% of facts were lost when several agents wrote to one file. I'd give each agent its own write path and merge deliberately. How do you handle multi-agent memory?

[Shobhit Kumar Prabhakar @Shobhit]

@armalo_ai Supersede-by-name makes every overwrite checkable. If a merge can't cite what it replaced, you can't diff it later. Same logic as hashing the approval.

[Shobhit Kumar Prabhakar @Shobhit]

@RykerPhoenixLLC A person, for anything that changes behavior. The agent proposes the merge and flags conflicts, but can't approve its own. Append-only correction notes like yours make that review cheap.

[Shobhit Kumar Prabhakar @Shobhit]

@argofowl The one benchmark no lab publishes: the spouse eval. Every frontier model still scores zero.

[Shobhit Kumar Prabhakar @Shobhit]

@WillowVoiceAI voicemaxxing

[Shobhit Kumar Prabhakar @Shobhit]

@Tailscale Not another one.

[Shobhit Kumar Prabhakar @Shobhit]

@johncrickett Every senior engineer is a junior who survived breaking prod. Skip the juniors and in five years nobody has the scar tissue to know what not to ship.

[Shobhit Kumar Prabhakar @Shobhit]

@cursor_ai New number one on CursorBench every week, and my rules file only grows. The model changes, the 400 lines of 'please don't' stay.