The pieces are slowly falling into place, and the chances of a concentrated take-off are reducing day by day. Bullish on a shared radically abundant future. {Sources (as shown with superscripts ⁰¹ ⁰² ⁰³) and further explanations are provided in the thread below} The

@EpochAIResearch @MatthewJBar You can even read it on @lesswrong by @boazbaraktcs - https://t.co/2f4KN0xGyh

@EpochAIResearch @MatthewJBar Q(t) is Quality - which corresponds to the fraction of economically useful tasks these workers can perform. If digital workers are deployed at all, I do not think frontier labs will make them generally available with a comprehensive set of capabilities. Even if there are 100k

@EpochAIResearch Again, @MatthewJBar says, in his article - Even under quite pessimistic assumptions (where the elasticity of substitution is 0.5), U.S. GDP could still double if remote labor (just 34% of total labour) is made 10x and became automated. If such an economic expansion occurs over

@EpochAIResearch did a back-of-the-envelope estimate that OpenAI can run approximately 7 million digital workers, which are digital instances of GPT-5. https://t.co/4nfNRpoJPy Since both OpenAI's compute "capabilities" and the compute "requirements" to run a digital worker at

I completely agree. However, three different and related aspects should be noted here: 1. Task Heterogeneity - Not all "tasks" are the same. If "AI automates half of the tasks" in an industry, those tasks are most likely the "easiest to automate." 2. Cost - GPT-5 Pro costs

Obviously yes, but the impact is much much harder to assess. https://t.co/1QNWXtG3zG

This is a very good non-technical post from Prof. Barak, CS at Harvard and MTS at OpenAI. However, the growth from automation is much more nuanced than what is detailed in Barak's post. Here, @MatthewJBar explains⁰¹ the nuances: A brief primer on the elasticity of

This is probably one of the most impactful wisdom bits you'll need to understand about the future trajectories of Frontier Intelligence development. This is a weak mode of the Recursive Self-Improvement loop. But to achieve this, you need to have solid capabilities starting

There are many ways in which closed frontier labs can help advance humanity, apart from open-sourcing their STEM breakthroughs & knowledge. One of the most effective ways is using AI in the service of the Open Source Ecosystem. If there were an option to choose between releasing

Even @windsurf's SWE-1.5 is priced similar to 2.5 Pro & GPT-5 (medium reasoning) and 2X the price of Kimi-K2 and Qwen 480B Coder. I'm glad that these labs are on the way to become (or already) profitable. One cannot sustain loss making businesses for long term and I don't want

Assuming it's a fine-tune of Qwen 3 480B Coder or a new model as large as Qwen 3 480B Coder, these output prices are 5Xed! A 500% increase in output pricing will give them insane margins! Who is saying start-ups and frontier labs are burning VC money and serving models at a

And even on the Remote Labor Index (RLI), Sonnet 4.5 is above (a little bit) every other model. Manus substantially uses Sonnet 4.5 as its underlying model. https://t.co/37yrIY3okX

I was writing this tweet when they posted this! lol😆 https://t.co/QvYq8zCOLt Congrats on moving another step in the value chain, @cognition @windsurf team, and a much bigger congrats for serving this on @cerebras! 👏🏽👏🏽 @jacobtpl @connorffogarty @GoTypescript @theomarcu

This is the fifth or sixth time I've used this beautiful, simple three-line piece of deep wisdom. Sorry for tagging you so many times, Han. I always try to reference the source as much as possible. 😅 https://t.co/23dBCZksK1

It is great to see @cursor_ai, @cognition's @windsurf (with SWE-grep), @ManusAI (with Manus 1.5), @Replit (Agent 3), and many other large startups moving backward in the value chain to directly compete with the "model providers." @midjourney, @runwayml, @elevenlabsio,

TSMC is not the threat to free markets -- because its primary purpose is to serve as Taiwan's "silicon shield," rather than attempting to consume every layer before and after it in the stack. To date, it has not even hinted at moving either forward or backward in the value chain.

The Automation Machine is coming - again! FWIW, Claude models have been nearing human performance in the "economically valuable tasks" over the past six months. While the world is much more complex than these benchmarks suggest, they are the best proxies we have for studying

Even the additional share warranties do not vest automatically; they have to be bought and paid for by the OpenAI Foundation. All is not lost - since the OpenAI Foundation board has the sole governance and managerial rights (in addition to 26% stake) over the OpenAI Group, and

Grok seems to be sane & truthful fwiw!

My previous post discusses why open-sourcing AGI systems will be far better than sharing a quarter of financial profits with the world - https://t.co/80xbUZR8wQ $130B valuation and future share in earnings is a very good amount... but, I continue to have very strong reservations