Mitchell Griesemer

[Mitchell Griesemer @MGriesemer]

@gnawbone_ @thdxr Direct quote from Dario on his recent blog backing up what you said. Not sure how Dax can come to his conclusion after reading this.

[Mitchell Griesemer @MGriesemer]

@jeremyphoward The Trump admin is so incompetent that we blame Anthropic because we expect so little of our government. This is like when people were arguing Zelenskyy should have been nicer to Vance at the Oval Office while the president and vice president go around spreading RUS misinfo.

[Mitchell Griesemer @MGriesemer]

@joseph_h_garvin @DevelopmentsAI It makes perfect sense for this good enough code to be of lower quality than the average Anthropic employee but still meet a quality threshold that is okay for deployment. That scenario would lead to more LOC and a productivity speed up.

[Mitchell Griesemer @MGriesemer]

@joseph_h_garvin @DevelopmentsAI At no point does the article claim Mythos or Opus 4.8 are AGI. In fact, the paper basically says the code it writes is “good” but worded in a way that means “good enough.” They’re just arguing that AI has reached a point where the code quality is good enough for internal use.

[Mitchell Griesemer @MGriesemer]

@Never_Hedgee @arakharazian @DKThomp @tryramp Same for Anthropic. This is a result of smarter models that are more efficient thinkers. Mythos used less tokens than GPT 5.5 on cyber security coding tasks while scoring better.

Cost per task will likely to continue to fall by roughly 300x in 2026 the same way it did in 2025.

[Mitchell Griesemer @MGriesemer]

@JakeKAllDay @scaling01 Dario said in an interview that he only thinks Anthropic is about 1-3 months ahead of other frontier labs so he agrees with you. Nobody is claiming the other labs can’t get there. People are just coping about Mythos and downplaying how good it is. GPT 5.5 isn’t as good.

[Mitchell Griesemer @MGriesemer]

@celestepoasts MIT did a METR like eval for non-coding tasks and still found a doubling time around 3.8 months.

https://t.co/YkVrNyuebo

[Mitchell Griesemer @MGriesemer]

@deanwball @croissanthology Then we’re in agreement and I should have been more clear. We should build 125% of the current grid in solar by 2040 on top of what we currently have. Only 25% more would be a massive failure.

My pushback was mostly related to your framing of solar.

[Mitchell Griesemer @MGriesemer]

@deanwball @croissanthology Dude come on. That’s not a “loser premise” if our time frame is 2040. Fusion, SMRs, geothermal, etc. might all be a big deal in the future but right now, solar and batteries. They turned energy production into a manufacturing problem. Amazing tech.

[Mitchell Griesemer @MGriesemer]

@EconBerger Zero evidence. I’m genuinely curious. My intuition is that if investment spending was 10% in both 1870 and 2026, it would be harder for an industry to reach 2% in 2026 because modern economies are more complex and have more profitable investment options.

[Mitchell Griesemer @MGriesemer]

@petergostev Oh come on.

Opus 4.1 to Opus 4.5 - 100ish days

Opus 4.5 to Opus 4.6 - 90ish days

Opus 4.6 to Mythos - 61 days

Even if each of these jumps follows a clear trend line, the timeline still suggests Mythos is an exceptional release.

[Mitchell Griesemer @MGriesemer]

@JulieBorowski I would argue that “rational” in economics just means predictable. If human action was random then it would be impossible to study it and make predictions. When a group of people have perfect information they will tend to make very similar and predictable choices.

[Mitchell Griesemer @MGriesemer]

@petergostev Really don’t think this is true for some tasks. Gemini 3 flash does an exceptional job at extracting data from PDF files. I suspect as models improve the cheaper options will get more advantageous for a growing number of tasks.

[Mitchell Griesemer @MGriesemer]

@CameronCorduroy Then teach us! I’m assuming the argument is that red states have lower GDP per capita, so convergence makes higher growth rates easier?

[Mitchell Griesemer @MGriesemer]

@emollick That’s the situation at my bank. We’re on the Teams plan with Claude but can’t get the enterprise people to email us so I’m pretty restricted on what I can use it for.

[Mitchell Griesemer @MGriesemer]

@austinc3301 Bingo!

[Mitchell Griesemer @MGriesemer]

@jasonfurman If you knew with 100% certainty that we get something like AGI or Dario’s “country of geniuses in a data center” what would your phase change estimate be then?

[Mitchell Griesemer @MGriesemer]

@ReggBrown75 @hamandcheese There’s been algorithmic progress that increases inference efficiency and Nvidia is going to release some game changing chips this year. Blackwell is already starting to make an impact but Vera Rubin will cut costs another ~90%. Expect Anthropic to be profitable by 2028

[Mitchell Griesemer @MGriesemer]

@AhmedAlNeaimy @claudeai “Bro, just finish that 1 GW of compute you purchased 24 months ago.”