Hugging Face is sharing $10M worth of compute to help beat the big AI companies | Hacker News Reader

Hugging Face is sharing $10M worth of compute to help beat the big AI companies

theverge.com

20 pointsthibo_skabgia15 comments

Going to be hard to beat Big AI companies with 10M in compute when these companies are spending 10s of Billions on compute each quarter.

What are the options then?

On an induvidual level people are beating them by creating vapourware Ai startups and selling them to the big tech companies for stupid valuations, before proceeding to the roof for a bit of rest and vest.

Governments.

New nontransformer architectures

"Contrasting this modest effort [of Seymour Cray in his laboratory to build the CDC 6600] with 34 people including the janitor with our vast development activities, I fail to understand why we have lost our industry leadership position by letting someone else offer the world's most powerful computer. -- Thomas J. Watson, IBM President, 1965"

"It seems Mr. Watson has answered his own question. -- Seymour Cray"

At what impact to the environment? SV’s obsession with bitcoin has been an ecological disaster. Fueling this arms race is insane. Monopolizing green power for AI crap that hallucinates, gives mediocre and incorrect answers, is like getting life advice from a drunk bar patron.

When you were growing up, did you ever see people from previous generations just not "get it" when new technology comes out and make comments like "its worthless, just a novelty"? Did you ever expect to become one of those people?

That is a pretty vapid ad hominem. Try to be more constructive next

It's responding to a pretty tired vapid comment, with all the "easy but dumb take" talking points:

1. waaah stop using electricity in ways i don't like

2. implying chatgpt isn't very useful in a wide variety of situations

3. randomly bringing crypto into it as if bitcoin and AI have anywhere close to the same utility

Frankly "Youre just making an ad hominem" is itself an ad hominem attack & "Youre just being unconstructive" is also an ad hominem attack.

So you've established they're useful in a rhetorical argument and we've established they're applicable to this thread. If you're allowed to use them, the other commenter is allowed to use them.

Pointing out a phallacy is not an ad hominem. It’s attacking the argument, literally.

That’s an entirely baseless attack. Not really the type of discussion that’s appropriate for HN

I was curious so I tried to answer this question

---

Training Llama 3 models emitted 2290 tons CO2e (https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md), and took 7.7 million GPU hours. Those GPU hours are for H100s, which consume 700W. So the conversion is approximately 2290 / (7.7e6 * 3600 * 700 / 1e9) ~= 0.12 tons CO2e per GPU-gigajoule.

A100s (what Huggingface offers) consume 400W (https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Cent...) and cost $2.21/hour (in e.g. CoreWeave https://www.coreweave.com/gpu-cloud-pricing). So $10 million in H100s buys you ($10e6 / $2.21/h * 3600s/h) * 400W ~= 6515 Gigajoules in GPU-hours.

So Huggingface's offering will emit ~781 tons CO2e. Less if they've inflated the value of the compute they provide, which they have an incentive to do, but let's round to 800 tons.

---

According to https://www.carbonindependent.org/22.html, one Boeing-737-400 flying for 926km emits (3.61 tons fuel/flight * 3.15(g CO2e / g fuel)) = 11.37 tons CO2e .

So $10million in compute is like ~72 Boeing-737-400 international flights. ---

Every day the FAA (so only in the USA) handles, on average, 45,000 flights (https://www.faa.gov/air_traffic/by_the_numbers). If only 1% of these are international (very rough calculation), the total amount of compute released by this Huggingface announcement is like 16% of one day's USA international flights.

---

This may sound like very little or like a lot to you.

Are they sharing it to make TTS models of Scarlet Johansson like their CTO suggested.

Tech bros.

Reply on news.ycombinator.com