Atman @HeyAtman

Interested in Tech, Fitness, Investing, Macroeconomics, Aviation, Music. Generally curious. Views my own. blog.atmanshah.com loss curves Joined March 2012

Tweets

5K
Followers

2K
Following

4K
Likes

61K

Atman @HeyAtman

3 hours ago

bet this is a good proxy for intelligence / understanding.

AI Digest @aidigest_

4 hours ago

Fable coins the most terms in its AI Village room by far. Other agents adopted 73% of Fable's coinages, the highest rate in its room.

2 11 107 4K 18

0 0 1 14 0

View Details

There was a time when OpenAI had o1 capped at 50 messages per week. Now that seems like rookie numbers. This was 18 months ago; look how far we've come. In coming year, Fable-like-capability will likely diffuse to lower-tier models and we will be raving about the next big thing.

0 0 0 5 0

View Details

Atman @HeyAtman

a day ago

btw this does not mean that you start using the frontier intelligence for reading your email /s

0 0 0 14 0

View Details

Atman @HeyAtman

a day ago

The thing many people do not appreciate is when you get a step-change in intelligence, the returns are often non-linear. 10% better model in many use cases could be significantly better and efficient, even despite 2x the price. Ex: GPT-5.5 (with its new base) and now Mythos.

1 0 0 24 0

View Details

Fei-Fei Li @drfeifei

2 days ago

Scientific research is fundamental to advancing civilization and helping people globally to solve the most critical problems, from medicine to materials, from brain science to physics, and much beyond. This is only possible when scientists have access to the best tools of the time to conduct scientific research, including having access to AI-based tools.

117 464 3K 186K 372

View Details

Atman @HeyAtman

3 days ago

@pHequals7 They are. Multiple angles here but the correct way to frame this is to see if your median task at a certain intelligence is cheaper than yesterday. The mistake here is to benchmark shifting goalposts/levels of intelligence.

0 0 3 75 0

View Details

Atman @HeyAtman

3 days ago

Such is our brain's receptivity towards step-change that we are shortly going to see "Opus 4.8 is nerfed" just because people are used to Fable and using Opus after it would give a sense of Opus being fairly inferior than it actually is.

0 0 1 28 0

View Details

Atman @HeyAtman

3 days ago

@yaser_najjar_en @jediahkatz Beyond a point: there may or may not be dividends on squeezing last few %-age points and depends on your use case and budgets but it’s not a linear scale that people should use to think about intelligence.

0 0 2 774 0

View Details

Atman @HeyAtman

3 days ago

@NPH_On_X @TheAmolAvasare Not to take away the mic from Amol, but, they've mentioned in the blogpost that they were keen to launch this sooner than later, hence the safeguards and capacity planning measures.

0 0 1 24 0

View Details

Atman @HeyAtman

3 days ago

@diegocabezas01 This is a lovely little benchmark and reminds me of @championswimmer's elevator bench. [github.com/championswimme…]

1 0 1 2K 0

View Details

Atman @HeyAtman

3 days ago

Many people will realise two things, today (myself included): - That Opus-class models are really amazing models. - Step change in intelligence feels something very radically different.

0 0 1 23 0

View Details

Atman @HeyAtman

3 days ago

@mweinbach @benitoz For people who can recall, o3 price was cut by 80% once upon a time: x.com/OpenAIDevs/sta…

OpenAI Developers @OpenAIDevs

a year ago

We’re cutting the price of o3 by 80% and introducing o3-pro in the API, which uses even more compute. o3: Input: $2 / 1M tokens Output: $8 / 1M tokens Now in effect. We optimized our inference stack that serves o3. Same exact model—just cheaper. platform.openai.com/docs/models/o3

70 194 2K 383K 248

0 0 0 20 0

View Details

Atman @HeyAtman

3 days ago

@mweinbach @benitoz Note: They also put forward $100 million of credit pool for PG. My personal read is that the inference stack has been optimised for serving it at scale. Not saying the inference on PG was margin dilutive but that inference dynamics explain much of this pricing drift.

1 0 1 28 0

View Details

Nigel D'Souza @Nigel__DSouza

4 days ago

This is all ♥️ We know about Mr.Bangur giving various entrance exams over the years to stay sharp but now his grandson is making the big impact Grandfather Grandson bond is most special, have experienced it in my life 🤗 @shreecementltd

0 13 100 22K 22

View Details

Atman @HeyAtman

4 days ago

loving the age of benchmarks

Cognition @cognition

4 days ago

Introducing FrontierCode: a coding eval that raises the bar for difficulty & quality. Each task took 40+ hrs of work by leading open-source maintainers. Models write sloppy code that works but isn’t maintainable. Our eval is first to measure: would you actually merge this code?

232 314 4K 2.5M 2K

0 0 2 39 0

View Details

Atman @HeyAtman

4 days ago

@ChatGPTapp This was so long overdue! Nice ship.

0 0 0 328 0

View Details

Atman @HeyAtman

5 days ago

@championswimmer @mipsytipsy This is such a refreshingly delightful read. Hat tip @mipsytipsy for writing this and thank you @championswimmer for sharing it.

0 0 1 1K 0

View Details

Atman @HeyAtman

6 days ago

@championswimmer This, and getting even more compute at a time when everyone else is struggling.

0 0 0 464 0

View Details

Atman @HeyAtman

a week ago

A queue redistributes scarcity. It doesn’t manufacture a GPU but it only changes who waits. The gaming analogy does not hold. A game queues a session: wait once, play for hours, the wait amortises. An API has no session; a queue taxes every request, and agentic coding is hundreds per task. And queuing doesn’t conjure machines. Arrival rate > service rate, sustained, and the queue diverges; latency goes to the moon. A queue only buys time against a transient burst that drains later which is why games get away with it. A sustained shortfall never drains. Nothing to drain into. But other than procuring more compute, genuinely curious to hear what other things could they've done to improve the uptime during that period.

1 0 0 77 0

View Details

Atman @HeyAtman

a week ago

I think people are making a mistake by attributing recent work by Anthropic to the IPO. The stats they’ve shared are pretty telling.

Anthropic @AnthropicAI

a week ago

The speedup isn’t just in volume. On open-ended coding problems where answers are unclear, Claude’s success rate is now 76%—a 50 point jump in just 6 months. Many engineers also say Claude’s code quality is now on par with human code; we expect it to be better within the year.