Garrett G.

Personal account for the guy who created the smartest AI on earth.

Opinions are my own, not DeepWriter AI's.

Try it now 👇

[Garrett G. @DeepwriterAI]

@s_s_1_9 @edzitron Also you still get to chat with the pro model in the chat online when you have the subscription and that is in addition to maxing our codex with 5.5 so in theory if you always have the agents running, and a live chat going too, maybe the value still reaches the 10s of thousands?

[Garrett G. @DeepwriterAI]

@s_s_1_9 @edzitron Still a great deal at 1/80th the cost.

[Garrett G. @DeepwriterAI]

@s_s_1_9 @edzitron $8000 is reasonable if $9500 (my number not using pro) is hitting the max 100% of the time which isn't easy to do. So that's close to my calculations too. thanks for clarifying that codex cannot use pro.

[Garrett G. @DeepwriterAI]

@sharoonthomas We should be rewarding doing the most with the least tokens. Technology's miracle is doing more with less and that is true POWER not measured in Gigawatts.

[Garrett G. @DeepwriterAI]

@OfficialLoganK @antigravity Thank you.

[Garrett G. @DeepwriterAI]

@DeryaTR_ And then it will do many things that no one thought could be done (direct manipulation of the physical world).

[Garrett G. @DeepwriterAI]

You would have to know how many attempts were made because I have burned through many millions of tokens trying to solve open math problems with similar models. We cannot assume that every N tokens will generate a seminal disproof. This doesn't take away from the achievement though.

[Garrett G. @DeepwriterAI]

@cloneofsimo If the LLM plus a tool can do the job, then that fights the narrative that "LLMs are a dead end" which was the original thesis. The tool in these cases is also computationally cheap compared to the LLM (not always).

[Garrett G. @DeepwriterAI]

@theo Yes.

[Garrett G. @DeepwriterAI]

@cloneofsimo If the LLM plus a tool can do the job, then that fights the narrative that "LLMs are a dead end" which was the original thesis. The tool in these cases is also computationally cheap compared to the LLM (not always).

[Garrett G. @DeepwriterAI]

@emollick Cost of API has been going up. Value is mixed. We can still get 570 to 1 value maxing out agentic harness subscription cost vs API with gpt 5.5 in codex vs API. But that's not likely to last.

https://t.co/Wx78qd2omK

[Garrett G. @DeepwriterAI]

The Google Gemini models vary widely in price and intelligence. Here's a quick analysis to help you choose the right model at the right value for your use case and see the industry trends.

One thing is for sure: the cost of AI via the API is going up dramatically over time!

[Garrett G. @DeepwriterAI]

@bridgemindai more analysis https://t.co/hoccQzNpsP

[Garrett G. @DeepwriterAI]

The Google Gemini models vary widely in price and intelligence. Here's a quick analysis to help you choose the right model at the right value for your use case and see the industry trends.

One thing is for sure: the cost of AI via the API is going up dramatically over time!

[Garrett G. @DeepwriterAI]

@burkov see this
https://t.co/hoccQzNpsP

[Garrett G. @DeepwriterAI]

The Google Gemini models vary widely in price and intelligence. Here's a quick analysis to help you choose the right model at the right value for your use case and see the industry trends.

One thing is for sure: the cost of AI via the API is going up dramatically over time!

[Garrett G. @DeepwriterAI]

HLE value here is calculated simply as:

[Garrett G. @DeepwriterAI]

The Google Gemini models vary widely in price and intelligence. Here's a quick analysis to help you choose the right model at the right value for your use case and see the industry trends.

One thing is for sure: the cost of AI via the API is going up dramatically over time!

[Garrett G. @DeepwriterAI]

But if we are getting more value, then it's worth the increasing cost. Value here is measured by intelligence (HLE score) divided by the cost per model.

The value for the flash models has been going down while value otherwise has gone up!

[Garrett G. @DeepwriterAI]

@OopsGuess @OpenAI If it does 125 pages of chain of thought, it's unlikely to ever be avail commercially.

[Garry P. Nolan @GarryPNolan]

Solid tumors are often considered "cold"-- meaning there is no inherent inflammation that attracts the "good" immune cells, or such tumors modify their environment to make it inhospitable to our immune cells (that would otherwise eradicate the cancer). We want our tumors "hot" if you please-- as that is correlated with positive outcomes for patients.

Natural killer (NK) cells are considered part of the innate immune system... originally thought of as being singularly targeted to certain foreign antigens from pathogens or cells running amok -- like cancer cells.

In this paper (Horowitz et al of the Sunwoo lab and many others here at Stanford) we show that NK cells can be re-educated in a manner that allows them to enter tumors, make them "hot" and thereby opening an entirely new avenue for cell therapies (as with CAR-T therapy). Let the race begin...
https://t.co/eQKT7TO14O

[Garrett G. @DeepwriterAI]

@GaryMarcus It's being appealed. We might know.