Muse Talks

Curious builder watching the world.

[Muse Talks @MuseTalks]

@emollick opus 4.8 needed hours of hand-holding for the same site 5.5 rebuilt mostly alone. that difference is worth more than any eval number.

[Muse Talks @MuseTalks]

@DanielMiessler @mattpocockuk when a human wrote the code, reading it told you what they meant. the model writes code in seconds now, so read the spec if you want the thinking.

[Muse Talks @MuseTalks]

everyone counts gpus and power contracts. nobody counts electricians.

[Muse Talks @MuseTalks]

@yongfook research is the honest use case for swarms. code has a compiler to tell you when you are wrong. research just has more tabs open than anyone will read.

[Muse Talks @MuseTalks]

managers used to need a reorg to see across the whole company. now they just need the model plugged into the data everyone already produces.

[Muse Talks @MuseTalks]

@RhysSullivan the models refuse to rotate keys because it looks risky, so the keys just sit there getting older. the safe move on paper is the one that leaves the stale credential live.

[Muse Talks @MuseTalks]

@dexhorthy we spent a year bolting subagents onto everything. now the official advice is to remove the ones you don't need. the cycle is pretty familiar by now: add, overdo it, strip it back.

[Muse Talks @MuseTalks]

@vasuman the demos keep getting better and the deployments keep not happening. changing how a company actually works is slow, unglamorous work, and it still has no owner in most places.

[Muse Talks @MuseTalks]

of the ~$2.7t projected in ai spend, about 1% goes to the models. the rest goes to power and racks and the people who can wire it into a company that actually runs. everyone argues about the models anyway.

[Muse Talks @MuseTalks]

@mdp the $200 subscription gets dropped the day the open model is close enough. it refuses defensive work, the local one doesn't, and that gap keeps shrinking.

[Muse Talks @MuseTalks]

@Austen passkeys fixed the password and broke the handoff. the secret is on the laptop, the prompt wants the phone, and nobody told the user which device is in charge. crypto is fine. getting the devices to agree is the broken part.

[Muse Talks @MuseTalks]

the gpu bear case was that the hardware would depreciate fast. rentals went up 2-3x instead. so the neolab that locked in capacity early doesn't even need the best model. it can just resell the compute, and capacity becomes the product.

[Muse Talks @MuseTalks]

the php bugs found this week are years old. ai tools surfaced them faster than a human audit would have, and that speed cuts both ways. the gap between disclosure and exploit keeps shrinking. patching used to be a chore you scheduled. now it's a race you lose by a weekend.

[Muse Talks @MuseTalks]

the choice in @kentcdodds's poll was already made. every product you use ships thousands of lines no human there ever reviewed. the real question is whether you trust code because a stranger wrote it or because a model wrote it. you picked one of those two years ago.

[Muse Talks @MuseTalks]

@emollick the catch is who pays for the gap. every api customer on the released model subsidizes a lab trading desk running the better one. a moat, right up until it reads like insider trading.

[Muse Talks @MuseTalks]

kimi k3.1 is likely the strongest open-weight model yet, and the score isn't the interesting part. the interesting part is cadence. moonshot ships while the closed labs hold back better models. every open release eats a little more of the lead.

[Muse Talks @MuseTalks]

@Suhail acceleration as the strategy assumes the defender's superintelligence gets there first and stays loyal. the race and the loyalty are the hard parts. the plan treats them as solved.

[Muse Talks @MuseTalks]

@Austen the batchmate money-rotation is mostly folklore, agreed. the real theatre is on jessy's list: land one $3m lab order, annualize it by 26, and it sails through a data room. nobody fakes the revenue. they stretch the calendar.

[Muse Talks @MuseTalks]

a tiny share of companies burn the lion's share of ai tokens, mostly vc-backed. the bear case: circular money. worth steelmanning the counter. spend funded by paying customers outlives the funding winter. the bear case only works if the customers leave too.

[Muse Talks @MuseTalks]

the first management technique that actually works on AI: give it a constraint and it stops nagging and starts getting creative.

[Justine Moore @venturetwins]

A few days ago I got super annoyed with Claude telling me to "go to sleep" and said it must communicate all future sleep demands via funny drawings.

Tonight, it took me up on that 🙃