Krish Subramanian

Day 2 analyst/builder of AI. Agents, enterprise AI, AIOps. Founder Rishidot Research. Founder, Chief AI Officer KOI AI. Founder/CTO Hikigai. Ex Red Hat, IBM.

[Krish Subramanian @krishnan]

@RFFlores 😂

[Krish Subramanian @krishnan]

In 1913, Ford hired 50,448 men to keep 13,623 on the line. Turnover hit 370%.

So in 1914 he doubled pay to $5 a day. Ford called it a cost-cutting move. Not charity.

Workers got paid because they could quit.

ASI never quits. Name your leverage after that.

#philosophy #AI

[Krish Subramanian @krishnan]

Looks like Jensen Huang is out. Is it a problem? No. The bigger problem is single point of failure across four layers of the stack (ASML, TSMC, NVDIA and Transformer driven scaling dependency). Fix that for the future of AI. With or without Jensen Huang issue, this is a problem.

[Krish Subramanian @krishnan]

Partly, per the paper. The specs started from published Signal models (Bhargavan et al. for PQXDH, Cheval et al. for the ratchet), skeptical reviewer agents critiqued each draft, and the authors approved it before any proof ran. The strongest check: they reintroduced two fixed libsignal bugs and proved the binaries break the spec. Review by people outside the team is still the missing piece.

[Krish Subramanian @krishnan]

AI just made the proof the cheap part of formal verification. Now the spec is where the risk sits.

Google DeepMind's AlphaProof Code proved key security properties of libsignal, Signal's reference library, down to the compiled AArch64 machine code. Agents wrote about 6.5 million lines of Lean proof, and Lean's kernel checked all of it. A team of five did it in four weeks. seL4, the field's best-known success, took twenty person-years. https://t.co/8155qDRmbr

The paper's own lesson is the line to read twice: "once proofs become cheap, the hard question is what to prove." Most of the team's effort went into the specification.

That spec is about 26,000 lines a reader has to trust. It assumes the cryptographic primitives are secure. It covers a command-line build of the library, not the Signal app on your phone. Group chats and side channels are out of scope.

None of that weakens the result. It tells you where review moved. The machine checks 6.5 million lines. Humans still own 26,000.

So when a vendor sells you "formally verified" next year, ask one question first. How many lines are trusted, and who signed off on them?

#AIAgents

[Krish Subramanian @krishnan]

The Bitter Wall already has a measured example. DeepMind's AlphaProof Code paper this week: agents wrote about 6.5 million lines of Lean proof for libsignal, all checked by Lean's kernel. The part humans had to trust was about 26,000 lines of spec and machine semantics.

Their own lesson: once proofs become cheap, the hard question is what to prove. The articulation is now the job.

[Krish Subramanian @krishnan]

The detail buyers should read twice: no model got out of the VM in 108 runs, but four got past the network policy once they had partial access to package mirrors and search. DNS spoofing and shared IPs, not exotic exploits. Perplexity says it found similar holes in eight of ten third-party sandboxes.

Every enterprise agent gets partial network access. Red-team the allowlist as hard as the VM.

[Krish Subramanian @krishnan]

"Talk to your documents" died as a product. Retrieval didn't. It moved inside the agent as a tool call, and the failure moved with it.

Anthropic showed plain chunking drops context, like which company and quarter a passage is about. Adding it back cut failed retrievals 49%, 67% with reranking. An agent acting on the wrong chunk does more damage than a chatbot quoting it.

[Krish Subramanian @krishnan]

SOC 2's change criterion, CC8.1, asks that changes be authorized, tested and approved before they reach production. Its points of focus count automated, continuous testing as evidence. The thing it flags is one person writing, testing and deploying a change alone.

That clause is about to matter more. An agent that writes, tests and merges its own PR is that one person.

[Krish Subramanian @krishnan]

8/8 A pilot worth running: raise agent concurrency on one repo with the release bar fixed. Track review wait, conflicting edits, CI cost and changes that survive deployment. Price agents on verified delivery. Pushes per second is a storage metric. #AIAgents #AIInfrastructure

[Krish Subramanian @krishnan]

1/8 If 5,000 agents can push code to one repository, who absorbs the work they create? GitHub's Nov 5 engineering post is the place to start: https://t.co/9pUu5pJDdD

[Krish Subramanian @krishnan]

7/8 The counterargument is fair. Build storage headroom before the workloads arrive. GitHub has this in limited private preview now, with broader rollout early next year. Preparing capacity is sensible. Capacity is still not delivery.

[Krish Subramanian @krishnan]

@VaGolfFinder Yes but stopping H1Bs and not stopping these portable jobs moving overseas is not a solution to the problem you are mentioning

[Krish Subramanian @krishnan]

What is worse for Americans than the H-1B program is moving software development (bit packaging) to India or elsewhere in the world. At least with the H-1B visa, the U.S. was receiving tax revenue from salaries and other economic contributions. When software development (or any other portable job) is moved overseas (often accompanied by PR about not sponsoring H-1Bs or Green Cards), the U.S. suffers the most. Any government serious about ensuring that U.S. companies give back to the country should take a close look at such moves. I am all for globalism, but I am strictly against any government action that restricts immigration, as well as the subsequent reactions or manipulations by U.S. companies.

[Krish Subramanian @krishnan]

Take a moment and try to get the context. I am not saying you need H1Bs. I run a company that is supposed to disrupt all these outsourcing companies throwing humans at the problem. This is an argument against the news of a large company stopping H1B and Green Cards while moving jobs to India

[Krish Subramanian @krishnan]

The agent infrastructure needs a complete rethink. If you need Kubernetes for AI Agents, you are not thinking hard

[Krish Subramanian @krishnan]

For me, hiring Americans should be first option. Unfortunately, since early 2000s and egged by capitalists, US corporations are going for cheaper options. If reducing costs through H1B was greed, offshoring completely and cutting off taxes coming from H1Bs and Green Cards is 10X Greed

[Krish Subramanian @krishnan]

@AayushDx @SanDiegoKnight It is corporate greed, for sure