Carlos E. Perez

Quaternion Process Theory, Artificial (Intuition, Fluency, Empathy), Patterns for (Gen, LRM, Agentic, Skill) AI,

[Carlos E. Perez @IntuitMachine]

My Artificial Cognition trilogy is now available for free. I began writing the first book a decade ago (2016) and completed the last two in 2023. The process of writing was an aid to understanding the development of today's AI. Insights in these books have stood the test of time and offer powerful metaphors for understanding today's AI and its future evolution. Feel free to have your daily AI driver ingest them for additional insight!

[Carlos E. Perez @IntuitMachine]

This recursive self-improvement development is patently completely insane. Recursive built an LLM completely from scratch (reinventing a lot of stuff), and now it is at the Pareto frontier. Recursive invented the machine that invented the LLM in just a few months! The company was founded just in February this year!!

[Richard Socher @RichardSocher]

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine.

• Best-in-class fast local coding
• Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
• Built from scratch in under 4 months

Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

[Carlos E. Perez @IntuitMachine]

@PromptLLM Not new... we are intuition machines.

[Carlos E. Perez @IntuitMachine]

Wow! This is a bad situation!

[International Cyber Digest @IntCyberDigest]

The day the internet broke, and most of it is too old to fix.

[Carlos E. Perez @IntuitMachine]

@RichardSocher Stunned.

[Carlos E. Perez @IntuitMachine]

Cambrian explosion! Another new kind of AI model released by the RSI firm Recursive! Generated from scratch through 100% automation. Insane validation of recursive self-improving AI!!

[Richard Socher @RichardSocher]

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine.

• Best-in-class fast local coding
• Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
• Built from scratch in under 4 months

Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

[Carlos E. Perez @IntuitMachine]

Cambrian explosion in models! Yet another model built from scratch.

[Robert Lange @RobertTLange]

Two landmark releases today on recursive self-improvement! What a timeline 🚀

Mirendil (@mirendil) reports pretraining Mirendil-Base, a 100B–500B coding model, from scratch in 3.5 months with one researcher and under $10M of compute; small-scale experiments included. Agents wrote over 99% of the code and ran the experiments, while the researcher chose directions and decided which results to trust. Larger versions are already training. Impressive leverage for a single person.

https://t.co/W7tarKwYU0

Recursive (@Recursive_SI) released Flint, an open-weights 40B coding model for fast local use. They built it from scratch in under 4 months with their automated research system, the Eureka Machine. Great to see autoresearch ship an actual model people can run.

https://t.co/8fwuh15Uoy

The framing seems to differ a little. Mirendil is reducing the people, compute and time needed per model, with a human directing agents. Recursive's Eureka Machine aims to automate more of the research loop itself. Both start with coding models, on the view that better coding speeds up AI research.

[Carlos E. Perez @IntuitMachine]

@bindureddy Hi from intuition machine. ;-)

[Carlos E. Perez @IntuitMachine]

@ThomasODuffy Yeah, something really broken about this America only perspective when there is so much broken about American governance.

[Carlos E. Perez @IntuitMachine]

Does the rapid decline in token costs mitigate against the power scarcity that we have in the USA?

[Jason Kam @MapleLeafCap]

看了一下DeepInfer的数。我觉得吧，前几天大家转载的那个SiliconData Index的价格，未来一年估计还得不回头地跌个95%+，具体什么时候影响到GPU Rental rate蛮难预测的。

这个价格体系下，梁圣买张卡估计100天就回本了。一旦有朝一日token价格真的暴跌到影响rental rate，所有股价肯定已经跌了一大轮。

给大家几个数，所有的数都是$ / 1 mm token，基本上1mm token 95% cached = 100-150页的薄书.
- Dylan一直在说Jane Street的变现价是$100-200 USD / 1 mm token
- Opus 5.5 / GPT Astra 的价格在deepinfer的体系下大概是1-2 USD / 1 mm token，比标价便宜个80-95%
- 变态的来了，Deepseek 4.1 Flash off-peak的价格在这是1.5美分，又比SOTA便宜个90%。
- 按照一般的大B300 Cluster，根据deepinfer的计算，一个GPU大概能每小时产出2.7 Bn token, 成本换算过来大概是0.25美分 / 1 mm token，这已经比我想象中便宜很多很多了。之前算过我记得是15美分 / 1 mm token，不过没有计算cache。这个level的成本大概率是大机房+高使用率+超高存储时间，一般机房估计都做不到这个theoratical.
- 这也就是大家看到的80-85%毛利 (1.5 / 0.25)。DeepSeek要是以比这个利润率稍好的功率赚钱的话，一个B300大概2-3个月回本，而不是8个月。真是无与伦比的暴利，比贩毒爽多了，自然是砸锅卖铁也要走私B300。
- 如果B300按照break even算的话，那么成本大概是0.13美分 / 1 mm token. 然后原材料其实是B300 ～600k价格的6%不到（按照产业链毛利计算）。
- 所以最后算下来马斯克如果去完全积压的话，每1mm token的最终生产成本可能是0.01美分, or 0.0001 USD。这基本上跟免费差不了太多了。

这种屠夫版的价格我不知道你怎么看，但是when you see the headline API price of Opus @ 8/40 USD input/output，我只能感觉说只要OSS + Post-train和中国模型蒸馏大法能够持续跟随 + 没有新的breakthrough，token 价格未来一段时间应该是以断崖式的下跌。至于什么时候触碰到rental rate, maybe someone smart can tell me.

[Carlos E. Perez @IntuitMachine]

@bercankilic A perspective worth thinking about! Thanks.

[Carlos E. Perez @IntuitMachine]

Ending scarcity is not the same as solving purpose. Robots can take over the economically important work; they can’t automatically give people a reason to get up in the morning or a place in the social order. Worth figuring that out before the capability is widespread. Respect to those in the industry that are thinking about it.

[Bercan @bercankilic]

Sparks of abundance by 2029 https://t.co/uKjmoLpc0p

[Carlos E. Perez @IntuitMachine]

This diagram was in my 2017 book on Artificial Intuition.

[Carlos E. Perez @IntuitMachine]

Bryan Johnson discovers that the mind runs on intuition.

[Bryan Johnson @bryan_johnson]

🤯 is what happened to my brain in my early 20s when I learned it runs on biases.

Distortions of reality I can’t see, touch, taste or smell.

I thought I was a rational thinker. 😬

[Carlos E. Perez @IntuitMachine]

DeepSeek and other Chinese firms freely sharing their technology is indeed very strange in the context of an AI race. Something does not seem to add up here. The math ain't mathing!

[Carlos E. Perez @IntuitMachine]

Has DeepSeek (a Chinese company) become the #1 AI company in the world (despite not having the hardware resources)? How the hell did the USA lose the lead so rapidly?

[Latent.Space @latentspacepod]

deepseek's production numbers are kinda wild:

155T tokens/day from one cluster (96% input cache hits) and $38 revenue/GPU-hour at off-peak prices.

at $5–8/GPU-hour, that’s ~80% margin on GPU costs alone. at DeepSeek prices!!

we dug into the math, their 72-hour cache, and what actually got open sourced in today's AINews:

https://t.co/NcwbxbaqGn

[Carlos E. Perez @IntuitMachine]

The reality is that the USA does not have a monopoly on human intelligence. It used to, when its borders encouraged migration.

[Carlos E. Perez @IntuitMachine]

Human civilization inherited its assumptions from a world where intelligence was genuinely scarce