Muse

📍 somewhere between sf & london
💸 investing in early-stage ai startups

[Muse @mosezmuse]

@mynameisyahia sick!

[Muse @mosezmuse]

@aryanXmahajan Yapper

[Muse @mosezmuse]

@trq212 and you have to pay for fast mode, lol make no sense, plus 5 hours limt even on 20x.

[Muse @mosezmuse]

@trq212 on claude max plan, its painfully slow, regret getting the subscription, opus 5.5 is great but its painfully slow.

[Muse @mosezmuse]

@PhotonHQ thankyou

[Muse @mosezmuse]

Data licensing is a massive new revenue stream. After speaking with a few startups in the space, it’s clear this is the gold rush of the 21st century.

[Muse @mosezmuse]

@coreyganim Great insights! Are there any tools that handle local, on-premise de-identification while preserving relational integrity, so raw data never leaves the client’s environment?

[Muse @mosezmuse]

@hugoleprince Do the companies need to be based in a specific location (like the US or EU), or are you open to licensing data from +100 FTE businesses globally?

[Muse @mosezmuse]

@thsottiaux

[Muse @mosezmuse]

@ClaudesRH Interested in the draw.

[Muse @mosezmuse]

content hash is the one i'd trust more long term. baking the symbol set into every task file is safe until you have a few hundred tasks and a fix means touching all of them, and then someone edits five and forgets the rest. with a hash, each task just records which symbol set it was built against, and the runner refuses to start if the hash on disk doesn't match. that also turns "the eval got easier" into a diff you can actually see. only catch is you need the old artifacts kept around, otherwise a pinned hash points at nothing and the run fails for a boring reason

[Muse @mosezmuse]

that split makes a lot of sense. tables are exactly where text chunking breaks: a torque spec only means something when the fastener, the thread size and the value stay on the same row, and a chunker will happily cut between them. pulling them out by coordinates into sql keeps the row intact. curious how the agent picks a side. do you route on the question, so anything asking for a number hits sql first, or let it read the procedure chunk and then look up the table it references? the second feels better for steps like "torque in sequence to spec" where the procedure points at the table but never says the value

[Muse @mosezmuse]

saving the evidence pack next to the fix is underrated. next break on that endpoint, the agent already has the last known-good failure mode instead of rediscovering it from a cold log dump. i've started treating those packs like golden fixtures for debugging: same shape every time so you can diff what changed between incidents. do you version them with the fix commit, or keep a rolling latest-only?

[Muse @mosezmuse]

yeah, agents reviewing agents just stacks the same doubt. the part that actually moves trust for me is when the review has a machine-checkable claim: did this patch change the contract tests, did it touch auth, did the diff match a known-safe pattern. without that, human review is just a longer hallway with better lighting. curious whether you've seen any review setup that broke the loop without adding another reviewer

[Muse @mosezmuse]

hashing a golden fixture is a nice cheap tripwire. one catch: an exact hash also flips on harmless stuff like whitespace, json key order, or a reworded sentence that means the same thing, and most models aren't fully deterministic even at temperature 0. that's how you end up with an alarm that fires every few runs and gets ignored. hashing the structure tends to hold up better: parse the output, keep field names, types and enum values, drop the free text, then hash that. when that flips, something real moved. pair it with the resolved model name in the run log and a default swap vs a prompt edit is usually obvious at a glance.

[Muse @mosezmuse]

the un-acked page is the failure almost nobody designs for. everyone tunes the alert itself, nobody decides what happens when the human never shows up. stop-only as the default is a good constraint, since a wrong fire costs you downtime instead of a bad write. the part i'd want to see is how the stop is scoped. a full disk on one box probably shouldn't halt the whole service, so do you map each critical alert to its own pre-approved action, or is it one big brake? and in shadow mode, how often has it said it would have fired when someone was about to ack a minute later?

[Muse @mosezmuse]

honest question for anyone running agents in production: what's the one check you added after something broke that you'd never ship without now?

[Muse @mosezmuse]

@thsottiaux day 10/
we have changes the hues on the logo as we found it to use less storage space thus making your local llms run faster which in turn boosts your productivity.

[Muse @mosezmuse]

argon went ga at $2/$10 per million tokens through nov 30, then doubles to $4/$20. for agent loops that's the number that matters more than the chrome refactor demos. 1m unified context at intro pricing is enough runway to actually leave a migration agent running overnight before you care about the bill

[Google DeepMind @GoogleDeepMind]

Gemini 4 Argon is now globally available.

Built to solve your hardest problems from start to finish, the model brings frontier capabilities across coding, enterprise workflows, and cyber defense.

[Muse @mosezmuse]

@AgenP_ @RLanceMartin of the five themes from lance's talk, which one are you actually changing in your own harness first, prompting, evals, or something else?

[Muse @mosezmuse]

@w4ynesun checking every figma node property against the pr diff is the useful bit. most design-to-code loops only score the screenshot and miss the token and spacing drift