Phil
@phill__1
Followers
1K
Following
4K
Statuses
2K
Currently working on Tech Support AI
H
Joined January 2019
@Tuxsoia @test_tm7873 I think you need a base model that has seen the entire internet. GPT-3.5 didn't know much at only 20B parameters, but modern post-training and RL on the huge GPT-4 base model would likely lead to similar performance as the current reasoning models.
1
0
1
@test_tm7873 According to Ilya, base models will not improve significantly, pre training is over, we will just add more and more RL to it. Let's see
1
0
1
@test_tm7873 Well, GPT-4.5 is the last one, so my prediction was just a little early. Until now, no one has beaten 3.6 Sonnet, which was released in early October. And if GPT-4.5 is like 10% better than 3.6 Sonnet, I would still argue this proves my point
2
0
1
@NickADobos If you want to change the reading speed or the voice, share the chat, get the eleven labs reader app and paste in the link. Works amazingly
1
0
8
@rileybrown_ai @perplexity_ai Get the Complexity Chrome extension. It has this feature plus a bunch more that make Perplexity at least 30% better to use: cplx[.]app Not affiliated, just a fan
0
0
1
@Presidentlin You are free to steal it :) Ideogram 2 is still the best image model ever made. Tried the new Imagen 3, not even close
1
0
3
@AGI_FromWalmart It's literally just custom instructions, if this is what they are hyping up pre Grok 3 launch, I don't have high hopes
0
0
2
@adonis_singh Wait, is it actually outputting code? Deep research never outputs full code implementations for me, potentially a prompting skill issue on my part tho
1
0
1
@UriGil3 Both can be true; Elon is a bully and likely not acting in good faith regarding open source AI. But OpenAI's non-profit charter is actually well thought through and beneficial for the general public instead of investors. That's why softbank wants to get rid of it
3
1
19
@SimplyDionysus @newyorkbadboy Yeah, but I am talking about the OpenAI one. Gemini Deep Research is not nearly as good, sadly
0
0
0
These open-r1 blog posts from Hugging Face are probably the best resource for understanding how reasoning models work, as well as staying up to date on all new LLM RL training discoveries from the OS community. Highly recommended!
0
0
6
It takes 1 to 3 deep research prompts to become better informed about any news topic than 99.5% of the population
0
0
2
@instalabsai @Joadys @NickADobos @grok Thanks! I have made a longer post with a bit more background on how the for profit transition works if you are interested:
What's Musk's Logic Behind the $97B OpenAI Offer? It's a pretty clever gambit: OpenAI Inc. (the non-profit) wants to convert to a for-profit company. But you cannot just turn a non-profit into a for-profit – that would be an incredible tax loophole. Instead, the new for-profit OpenAI company would need to pay out OpenAI Inc.'s technology and IP (likely in equity in the new for-profit company). The valuation is tricky since OpenAI Inc. is theoretically the sole controlling shareholder of the capped-profit subsidiary, OpenAI LP. But there have been some numbers floating around. Since the rumored SoftBank investment at a $260B valuation is dependent on the for-profit move, we're using the current ~$150B valuation. Control premiums in market transactions typically range between 20-30% of enterprise value; experts have predicted something around $30B-$40B. The key is, this valuation is ultimately signed off on by the California and Delaware Attorneys General. Now, if you want to block OpenAI from the for-profit transition, but have yet to be successful in court, what do you do? Make it as painful as possible. Elon Musk just gave regulators a perfect argument for why the non-profit should get $97B for selling their technology and IP. This would instantly make the non-profit the majority stakeholder at 62%. It's a clever move that throws a major wrench into the for-profit transition, potentially even stopping it dead in its tracks. Whether OpenAI accepts the offer or not (they won't), the mere existence of this valuation benchmark will be hard for regulators to ignore.
0
0
44