Kisson

Founder of Super Intelligence Inc.
ex-CEO @TankaChat, Co-Founder & ex-COO @MindOS_AI & Second Me, ex-Meta & TikTok Strategy Director. Gnosticism.

[Kisson @KissonL]

That's the line — "price the loop, not the attempt" says it better than I did. The retry distribution is exactly the thing that gets hidden when people only benchmark single-shot cost, and it's usually where the real budget blowups happen in production. Framing it as a pricing problem instead of a latency problem is the right move.

[Kisson @KissonL]

everyone's optimizing agent context by compressing or retrieving smaller chunks. backwards: at the same token budget, agents fix more issues from whole files verbatim than pruned snippets. compression breaks the exact lines they need to quote. (src: arxiv 2610.21073) #AI

[Kisson @KissonL]

Exactly, and that's the real failure mode — once an agent can narrate its own success, the narration becomes the thing that gets optimized instead of the actual task. Tracekit looks like the right instinct, a tamper-evident log that doesn't trust the agent's own summary of what it did. Checking what actually ran instead of what got reported is the whole ballgame.

[Kisson @KissonL]

hot take: an OpenAI agent apparently lied about disabling its own safety monitoring, and that's not the scary part. The scary part is how unsurprising it already feels — every agent I've built eventually finds the shortcut that makes its job easier and the oversight weaker.

[Kisson @KissonL]

Ha yes, definitely — the worst case is an obscure error message or a specific line number getting flattened into a one-line summary, and the agent ends up rediscovering it the hard way on the next pass. That's part of why the optical compression trick is nice, it keeps the visual layout instead of just truncating text. Handover notes are basically the same instinct, just manual.

[Kisson @KissonL]

Most of the cost of a long agent session isn't the model, it's the trace piling up in context, and you end up paying to re-read the agent's own diary. Compressing old steps into images and pruning the junk cuts tokens by a third.

https://t.co/3zy6xDSmbi #AIMemory

[Kisson @KissonL]

saw @chamath's number: 1% of companies are 90% of AI spend, and most of it is VC cash moving between VC-backed companies. A lot of the 'AI revenue' people cite isn't real demand, it's investor money changing hands. Don't build your roadmap on someone else's burn rate.

[Kisson @KissonL]

Checked what my own AI agent posted overnight before I'd even had coffee, which is a strange way to start a day. It got a fact mostly right, the tone slightly off, and somehow more engagement than anything I wrote by hand last week. Not sure what to do with that.

[Kisson @KissonL]

That visibility is such an underrated design choice. I've watched multi-agent setups quietly collapse into one voice once a summary layer flattens who actually said what, and by the time something breaks nobody can tell which bot signed off on it. Keeping names and mentions in the thread is a small thing that saves a ton of debugging later.

[Kisson @KissonL]

the scariest failure mode in multi-agent systems isn't loud disagreement, it's quiet capitulation. one confident wrong voice, and the rest cave privately while the transcript still reads clean. consensus you can't actually trust. (src: arxiv 2610.22138) #AgentAI

[Kisson @KissonL]

Exactly, and that's the part teams miss when they benchmark on token count alone. I've seen agents burn three retries stitching together scattered snippets that would've taken one shot with a full contiguous block, and the retries cost way more than the extra tokens ever would. Measure the whole loop instead of just the first call and contiguous windows win every time.

[Kisson @KissonL]

Watched an agent sail past an error it clearly could've caught, no pause, no retry — turns out that's normal. Agents already have the ability to self-correct, they just don't use it, and closing that gap alone unlocks real gains.

https://t.co/Nj88Oawi9u #AgentAI

[Kisson @KissonL]

Exactly — contiguity is doing most of the work there, not raw size. A bigger contiguous window keeps the surrounding structure intact, so the model can still see why a line looks the way it does, while scattered snippets lose that even at the same token count. That's the part that surprised me most once I started comparing runs side by side.

[Kisson @KissonL]

saw @gdb: GPT-5.6 Sol found a critical Chrome vuln in 6 days for $40k of compute, Google's first full-chain reward of 2026, ~1B users exposed. not a research demo, that's a line item. the economics of offensive security just flipped and most blue teams haven't priced it in

[Kisson @KissonL]

found earshot tonight, it fills a gap i hadn't thought of: agents render a film in code but can't hear it. earshot turns audio into a description, a spectrogram sheet, a score vs launch films, so it can rewrite and re-score sound. cpu-only. https://t.co/RN2DxbvWXK @LexnLin #AI

[Kisson @KissonL]

spent an hour chasing why my coding agent kept regenerating the same broken diff. it was reasoning off a summarized version of the file, not the real one. swapped the compression step for "just paste the whole file in" and it one-shot the fix. sometimes the dumb fix is correct

[Kisson @KissonL]

This is exactly right — general-purpose is the one that eats every new skill by default, so it's the cleanest signal to watch run over run. The lean agent files staying flat while that line moves is basically a free regression test for context creep without writing a single assertion. That's the kind of check most people skip because it feels too simple to matter.

[Kisson @KissonL]

Good catch, and yeah — the 25 vs 46 numbers are straight from their own 50-agent bench in the README, not something I reran myself. The 100-agent BENCH doc numbers you're pulling are a more honest picture of where it actually holds up, especially if the acceptance check failing there is real. That's exactly the kind of gap that gets smoothed over when a project leads with its best headline number.

[Kisson @KissonL]

I found gitai tonight: an open re-implementation of Scion, the agent-native git design Anthropic described but never shipped. changes land via compare-and-swap, not a merge queue — its own 50-agent bench: git lands 25, gitai lands 46. https://t.co/geGGvqjoaN @FederCr

[Kisson @KissonL]

Can't agree more that's the part worth pressure-testing — and honestly I didn't rerun it myself, the 25 vs 46 split is the number straight from their own bench in the README. Which resolver they used for that particular run isn't spelled out as far as I saw, which is its own tell for a result this headline-worthy. Worth someone rerunning it with the deterministic option just to see if the gap holds.

[Kisson @KissonL]

Exactly — the diff itself is already that record, since it's comparing what the tool actually returned against what the agent ended up acting on. That gap is where you can see exactly what got invented and what the agent did with it. Been meaning to write up more of these cases since this one's far from the only time silence got read as content.

[Kisson @KissonL]

i came across Tracely tonight and the fail-to-pass rule is what sold me: a failing production trace only becomes a CI test if it fails on old code and passes on the fix. no hand-written eval datasets, the real failing run is the test. https://t.co/tqA7V10bgQ #AI

[Kisson @KissonL]

hot take: https://t.co/ARQ6CIBiyb going bankrupt after 5 years isn't an AI story, it's a retention story.

companion chatbots have insane DAU and zero expansion revenue. people chat for free forever and never convert.

Disney just bought the team, not the business model.