3 ms·
What a useless article. OpenAI will obviously do many things before "running out of cash" - 1/ Implement more aggressive advertising 2/ Stop training new model
by OsrsNeedsf2P 9mo ago
What a useless article. OpenAI will obviously do many things before "running out of cash" -
1/ Implement more aggressive advertising
2/ Stop training new models
3/ Raise more funding
- 3eb7988a1663 9mo ago4) Sell all of those RAM chips they bought
- viraptor 9mo agoThey can't... not in an affordable way anyway. Those are integrated into other hardware they have. They don't buy the ram sticks others can use.
- CamperBob2 9mo agoThe RAM doesn't exist. What they bought were RAM futures.
- trollbridge 9mo agoYou can sell this, although the value would plummet if they did.
- zb3 9mo agoVery good outcome.
- dcre 9mo agoAnd it’s just a restatement of the original, which isn’t any good either. Really mediocre.
- Esophagus4 9mo agoThat wasn't the conclusion of the article (I'm assuming you're reacting to the headline and haven't read the article). The author's prediction is that OpenAI will get bought. I know it sounds kind of crazy, but there actually is precedent for that: Reid Hoffman sold his AI startup after he realized there was no possible way for an AI startup to compete with the Googles of the world with $100B in cash and giant free cash flow machines given how capital intensive it is to build AI.[1] To me, it's not that outlandish to think that if OpenAI really does need to spend a ton of money to survive, they will probably have to either raise or get bought (or find that magic money tree by monetizing their business). Because right now, they don't have the cash flow to compete with Google. Could obviously change, but I think that's where the author is coming from. [1]https://www.eesel.ai/blog/inflection-ai https://www.eesel.ai/blog/inflection-ai
- emp17344 9mo agoAdvertising doesn’t mean free money. Aggressive advertising will almost certainly drive away users.
- pier25 9mo ago> 2/ Stop training new models Google et al won't stop training new models.