I think one underappreciated thing that's happening with ChatGPT is that it's reminding people of how joyful it is to do totally mundane searches, i.e. looking up recipes, and *not* get indundated with a million fucking ads.
@brandonsilverman
We're just getting started. The enshittification stage comes later.
@brandonsilverman Yes, it is nice to get real results, and even phrase your search in the form of a question.
@brandonsilverman still in the early part of the cycle where they try to get people excited. if/when they achieve lock-in any ads will be deeply insidious. see https://www.youtube.com/watch?v=5Zmh5hbgHFI
Truman Show Product Placement

YouTube
@brandonsilverman my vague feeling of optimism is that it may be pretty hard for the big players to get a stranglehold on AI tech, it seems like the barriers to entry are pretty low. My fears of Peter Thiel's Robotic Oppression Force are receding
@stephenjudkins @brandonsilverman srsly? do you realise the amount of resources you have to throw at an llm to produce acceptable results?

@mawhrin @brandonsilverman i've been following it fairly closely and the cost of training appears reasonably affordable and is dropping pretty rapidly. At cloud costs it's in the low $10Ms now, which wouldn't be too hard to raise. It's likely it will get a whole lot cheaper

https://blog.heim.xyz/palm-training-cost/

Estimating 🌴PaLM's training cost

How much did Google's 530B parameter model PaLM's training cost? Something around $9M to $23M.

blog.heim.xyz
@mawhrin @stephenjudkins @brandonsilverman https://simonwillison.net/2023/Mar/17/beat-chatgpt-in-a-browser/

the costs are dropping extremely rapidly, Alpaca 7B is roughly equal to GPT3 (not chatgpt's 3.5) quality, and an estimated training cost of 85k USD.

Obviously, gpt4 is wayyy more expensive, but alpaca proves that finetuning is cheap, and its now open source available. Lots of implementations also dont need the wide skills of gpt4, but instead need to be finetuned to specialize, so lower performance than gpt4 is fine
Could you train a ChatGPT-beating model for $85,000 and run it in a browser?

I think it’s now possible to train a large language model with similar functionality to GPT-3 for $85,000. And I think we might soon be able to run the resulting …

@laurenshof @mawhrin @brandonsilverman yes, and even the "expensive" training costs are well within the bounds of medium-sized investors. It's simply not a big enough moat to protect the big players. If someone can credibly spend $20M to dethrone MS/Google there will be investors lining up with their checkbooks
@brandonsilverman @mmasnick Also, even when you’re not getting ads, search engines have tuned themselves to be basically shopping engines, and so searches are overwhelmed by results from e-commerce sites.
@brandonsilverman if it can also just give me the f—ing recipe without the novella sites write in front of it solely for SEO, that’s totally worth it!
@brandonsilverman
Why do people still tolerate ads, is it an iOS thing?
@brandonsilverman
Oh, and if you think GPT doesn't have ads 🤔
@brandonsilverman I don't see that lasting long.
@brandonsilverman Ads yes... but also, there's magic to a reduction in drudgery. Ask it to do something that you *know* how to do, but would be annoying, and get it back instantly? Even if not perfect, it really feels like magic. Ask it to write some complex regex or spreadsheet formulae for you...
@ag I never said there wasn't lots of other magic as well. There clearly is.
@brandonsilverman fair! I guess ad blockers just mean that I’ve been hit with other magic more!

@brandonsilverman yes! I am guardedly hopeful that we can center human intention in a new way using these AIs that grok grammer & logic.

They can do what you mean and we should expect them to.

Maybe the best guard for now is to grow that cultural expectation.

@brandonsilverman Have you verified that the ingredients are correct? I’ve done a number of searches where biographical details are pulled from other related people, not the subject I was searching on. Sounded plausible until I did further non-AI research
@brandonsilverman They'll not be far behind. Perhaps even more subtle ...

@brandonsilverman

Oh, you just wait. You'll get the surprise of your life.

@brandonsilverman It's trivially simple to block most ads, so people still seeing ads must somehow like them.
@brandonsilverman okay, but what is the quality of recipe you get in response? Presumably it's an LLM's guess about what chicken Kiev is, statistically, which doesn't sound amazing.
@brandonsilverman "Couldn't agree more! Using ChatGPT (not bing)has reminded me of how amazing it feels to search for something mundane, like a simple recipe without being bombarded with adds or social media buttons, in a conversational way.
@brandonsilverman so... It's growing with novelty, and will probably die with ads.
@brandonsilverman BBC has a whole recipe collection completely free of ads. Also free of "in 1860 my great great great grandfather was born in a horse trough in the middle of a Siberian winter in the desert .." crap. https://www.bbc.co.uk/food/recipes
Recipes

BBC Food

BBC Food
@brandonsilverman maybe AI is not the solution to their problems, but ad blockers are? Reminds me of the old days when some people insisted on solving every problem with blockchains.