Language models hallucinate because their training and evaluation reward guessing over admitting uncertainty. Models are unable to say “I don’t Know” because they focus on accuracy. Guessing can improve scores but leads to more confident errors (hallucinations).

Some facts are unpredictable or unavailable, making errors inevitable. The solution might be roughly penalizing errors more than uncertainty and rewarding honest abstention.

https://openai.com/index/why-language-models-hallucinate

OpenAI just released how people are using ChatpGPT. Coding 4.2% We live in a bubble

https://openai.com/index/how-people-are-using-chatgpt, https://x.com/atelicinvest/status/1967941812078915614

Claude also released how people are using their API and Claude.ai. The 1P API is used at 97% for automation.

AP2 (Agentic Payment Protocol) and x402 just unlocked a new level for AI agents. Agent can now actually pay other agents and MCPs.
https://x.com/sundarpichai/status/1968013016181641492

https://x.com/mattturck/status/1966545601140531707

As mentioned in part 1 of this AI Socratic edition in September we're organizing 2 AI dinners, hope to see you there:

#AI Dinner

Sep 10th
Solana Skyline, NYC

Link: https://lu.ma/ai-dinner-13.0

#AI Dinner

Sep 17th
Frontier Tower, SF

Link: https://lu.ma/ai-dinner-13.2

#AI Engineer (limited to 500 attendees)

Nov 20 - 22, 2025
New York City

We totally recommend this event. Currently working on getting a group discount for our community and a discount code for our readers. In the meantime if money are not a problem for you, go ahead and sign up, it's one of the best AI event.

Tempo is a new blockchain specifically designed for stablecoin payments. It's launched by Stripe and developed by Paradigm, a well known crypto developer lab.

Tempo is designed to enable global payments, payroll, remittances, tokenized deposits, microtransactions, and ✨ A2A payments ✨.

Stablecoins are a big deal because they enable to fortify and distribute the USD without the need of intermediary. The new US regulation in fact are not only friendly to stablecoins but also pushing for it. Recently Paolo Ardoino, founder of Tether a stablecoin with $170B in market cap, met with the POTUS to discuss exactly this.

Salesforce, Amazon and many more predict that the agent economy will surpass the SaaS economy by 2030, which means autonomous agents will pay for services via cryptocurrency rails.

Tempo is designed for both humans and agents.

Interestingly in 2019 Zuck/FB tried launching a similar cryptocurrency called Libra then renamed Diem — I wrote about Libra when it came out: Analysis of Libra part 1 and part 2. Libra failed because of the strong regulations, but clearly today we live in different times.

Even the GTM strategy is very similar, capture as many strong partnership as possible to increase distribution.

I'll try to write an in depth analysis about Tempo since it touches my 2 most interesting topics: cryptocurrency and AI agents.

https://tempo.xyz

An Harvard study tracked 285,000 firms and 62 million, from 2015 to 2025 workers concluding that 1 senior + 3 juniors is now equivalent to 1 senior + Claude. The effect is strongest in wholesale and retail sectors, where junior hiring fell by about 40%.

This shift may have lasting effects on career mobility and wage inequality.

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555

https://x.com/kyle/status/1963368557347311929

https://x.com/wolflovesmelon/status/1963736263246389415

This new launch might single handedly put Apple back in the AI game. From the demo at the Apple Event it looks very simple, if you talk to someone foreigner and set the translation, the airpods dims the other person voice while you ear the translation. Simple and effective.

September is the beginning of fundraising season. While Devin was not as good a year ago, it was the first semi-autonomous coding agent, like Codex or Claude Code, that integrate with Github and your company tools, and AFAIK is still the only one that has a browser mode to test the changes it implements. So, congrats to Cognition for the successful raise. https://x.com/cognition/status/1965086655821525280

This Series C funding round, led by @ASMLcompany, fuels Mistral AI scientific research to keep pushing the frontier of AI to tackle the most critical technological challenges faced by strategic industries.

https://x.com/MistralAI/status/1965311339368444003

Great content 🤩 video with amazing visuals. This video uses the sliding block puzzle Klotski to explain state spaces in CS. Each puzzle move maps onto a graph, showing how empty spaces and simple rules create complex, chaotic structures. It highlights how CS provides “x-ray vision” into hidden patterns in life.

https://www.youtube.com/watch?v=YGLNyHd2w10&themeRefresh=1

Dwarkesh thinks AGI is further than we think because, as R. Sutton said, AI utility is very strongly dependent on its learning. LLMs today don't have continuous learning, AI unlike humans, can't learn and improve while employed on a job. Also compute scale will hit a wall, and new breakthrough will be necessary. Short video but to the point!

https://www.youtube.com/watch?si=D2v2ItV\_2K0N0IMA&t=90&v=nyvmYnz6EAg&feature=youtu.be