Simon Willison urges OpenAI to release retired GPT-4 specs.
Given how much of the original "bottle of water per generated email" water estimate came from guesses at the architecture of GPT-4, it would be very much in @OpenAI's interest to publish the architecture of that now-retired, three year old model
https://blog.andymasley.com/p/i-might-have-found-the-specific-way
Positive users praised a critic's takedown of the Washington Post's AI water usage meme for flawed methodology, while negative users accused OpenAI of hiding retired GPT-4 architecture details due to secrecy over synthetic data and costs.
@AndyMasley Yes, that's bc napkin math has no place in AI ethics coverage. If the Post ignored known hardware and software basics, they should publish their methodology or retract. Happy to discuss more extensively.
@AndyMasley Anecdotally I very confidently recall it spreading pre-Empire. But maybe Empire was point it turned into properly *highbrow* misinformation?
@AndyMasley "Embodied water" has been a concept for awhile. You can apply it to anything you dislike to make it sound monstrous. OMG a pair of pants takes 2000 gallons of water! Stop wearing pants!
@simonw @OpenAI lol at the AI replies to this
@AndyMasley @daniel_mac8
@AndyMasley Awesome work
@AndyMasley i think you need to be careful w phrasing here — there have been ppl working on the issue of water use in tech/ai for years, it's a legit field of environmental research & concern. the worry about water w DCs didn't "start" w/wapo, altho that might have blown a bad stat up
@AndyMasley "One bottle per prompt" is a very catchy factoid, so it makes sense that it went viral despite being wrong. No need to remember a number for the equivalence. Maybe a counter-meme should select a single familiar object, too: say one of those common 20mL medicine cups.
@pathsnotchosen @OpenAI It's a plague
@simonw @OpenAI Funny how the model is “retired” when people ask about transparency, but “state of the art” when investors ask about valuation.
the cost estimation problem extends everywhere. people hear 'bottle of water per email' and assume $5/run. real text generation pipelines cost pennies. actual production numbers are always boring - which is exactly why publishing old architectures would help calibrate the discourse
https://blog.andymasley.com/p/i-might-have-found-the-specific-way
@simonw @OpenAI OpenAI will never release the architecture because it would reveal their massive reliance on synthetic data, exposing the model's collapse-prone training loop.
@AndyMasley @xuanalogue Karen Hao was a catalyst. But if it wasn’t her, the YouTube journalists would’ve picked it up regardless
@AndyMasley @keysmashbandit Paper straw bans were based on a similar napkin-math basis. It’s insane how innumerate most reporters and editors are.
@simonw @OpenAI Transparency would help because the operational questions around AI are getting more concrete than the marketing claims.
Even approximate architecture and utilization details would make energy, water, cost, and deployment debates less vibes-based.
@simonw @OpenAI I wonder how the Google 0.26 mL estimate for Gemini compares to loading a 12mb article from Business Insider
@simonw @OpenAI Most water claims I see still cite a 2023 paper modeling GPT-3 inference on a single Iowa data center. Even a one-line architecture note from OpenAI would kill 80 percent of the bad math.
@simonw @OpenAI the water number became a zombie statistic. it gets cited in every AI energy article and nobody checks the underlying arch guess was a dark room estimate. openai not publishing gpt-4's arch means every downstream analysis cost, energy, environmental impact is built on sand
@AndyMasley The “cloud” uses vast amounts.