Sam Altman is starting to panic

Had to pull the lever many times to make this beauty!

You might like this take:

The strongest argument in this video isn't that LLMs are expensive It is that many companies accidentally used token consumption as a KPI. Inference is a cost. Business outcomes are the KPI. The teams that figure out "value per token" will win. The teams optimizing for "tokens per engineer" won't

funniest part is that Uber head exploding moment was rediscovering goodharts law. nice.

Love this guy 😂 He created his own niche. Commentary on the ai world with deep insights mixed with comedy.

@edzitron You’re genuinely becoming one of the sanest voices out there on AI and I appreciate that. But please, get out of my head.

If someone owned a magical knowledge machine, they would never let you use it for $20 a month.

ai budgets are about to become the new cloud bills. everyone wants agents until the invoice starts acting autonomous too.

idk like he's taking a pretty level-headed approach relative to others in the space i wouldn't really bet against this guy

Mo -- have you seen this? I think you'd find it very interesting (and frightening).

Parts of your arguments are true but also other parts are wrong. Claude Code Codex, Gemeni, and many other generative models can now implement a stable SaaS in hours, worst case, in days, and several iterations, with less than $500 in spending. Something that used to cost 100s of thousands, and months of development work and several developers.

I have regularly felt like I'm 'hovering over it like a disappointed Asian Dad' 🤣

Omg, “It wrote a spec so long that it deleted itself in favor of all mankind” 🫡🤣😂

I don't understand why these companies just wouldn't spring for the high-end subscription so that they're token budget is capped. I have done the high-end subscription for a couple of the different services and never really hit my token caps doing so.... Maybe I'm just incompetent

Cuck ran out of other people's ideas 20 years ago. He doesn't know what he wants to do with AI. All he knows is he's got to be in it with the other kool kids or his stock will get hit. He renamed his entire company to a product move that completely failed. "Meta" as a concept doesn't even exist anymore. Basically, a multi-$Billion name change. It rankles him that his entire business is based on something as sordid as scraping users' info for ad dollars but his every attempt to innovate his way out has gone off a cliff. Now, bereft of any means of increasing revenue with capex costs piling up, he's circling the wagons and cutting all the "talent" he blindly hired in the hope something magical would happen.

“Every time you ask for a tiny change it regenerates the entire fucking universe” nailed it

The bill came due. Those hyped valuations need backfilling. The analogies... The dealer gave us the taste of crack cocaine for free, then ramped it up The baby formula was free until the breast milk dried up And it's still a high-tech magic 8-ball, wrong as often as it's right.

Ok now, I'm sold, if we can have self deleting specs. after they reach some level unnecessary complexity, dump all our resources as a species in LLMs!

“ holding a press conference to talk about your vow-of-silence “ “ all of them moaning like it’s their first threesome with the babysitter “ - Gosh, Mo is becoming a Dave Chappelle!! Keep posting , Mo!!

Yo, Vibe keyboard 🤣

As long as this companies don't provide tokens refunds when the LLMs don't produce what is expected, you will be just burning money like in a casino.

Epic 😂

Measuring engineers by token usage is like measuring drivers by gas burned.. I can waste both and still end up at the same destination. But, why? People nowdays just seem stupid enough not to get that... 🙄

Your analogies have found you again, Mo! 💪🏻😂

i think he started to panic a while ago... you don't market dump your shitcoins OTC if you're not already panicking.

Absolute hilarity, bravo.

I hope the AI slop continues just so I can watch your roasts

Technical point: each token in the context doesn't get reprocessed from scratch. That's what the KV cache is for. As long as you don't invalidate the cache or wait for it to be deleted, additional queries are only as expensive as the additional tokens processed and generated.

I keep coming back to the same idea: AI is super to for the "blank page problem"; create when you have nothing. If your application is huge, then you have to try and box in the AIs focus onto a small blank window and let it fill in. Too large of a window and spaghetti!

“…at a volume so insane no human can keep up…” I concur

The context window is the killer Serious users are going local Say as you will

"the entire csuite is sitting there moaning, like it's their first 3some with the babysitter..." you are a master storyteller, sir.

Instant follow after that explanation. Very well said man. I couldn't agree more... and yes, I'm an older software guy trying to embrace this new tech as much as I can ;)

Generate, generate, generate.

Oh yeah. The entire centralized model's going to go tits up as soon as the enterprise crowd figures out that Chinese models exist.