Updated on Sep 8th

The top AI developments from Jul 15–Sep 15, debated Socratically over AI dinners in NY and SF. This is our 1st draft.

---

**Note by the Author** This blog post is HAND CRAFTED 🤚. any use of em dash (—) or delve is a conscious decision by the writer (me) to convey something or for pure personal pleasure — I love em dashes, I started using them after reading on writing well by William Zinsser. Enjoy!

---

Sign up to receive the mailing list!

#Upcoming Events

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: The Blockchain Designed for Payments

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

#Junior engineers are down 23% Senior engineers are up 14%

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

#Are we in the first inning of an anti-AI movement?

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

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

#Apple Airpods Live Translation

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.

#Cognition Raises 400m at 10.2B

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

#Mistral Raises $1.7B

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

#Graph Theory in State-Space

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: Two bottlenecks to AGI

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

#Full Sources List

Here the list of all the AI updates that couldn't make it to the headlines.

#AI Agents

• General Intelligence Company launches Cofounder an AI agent that works like a cofounder with your startup x.com/ndrewpignanelli/status/1965452670128587033

#AI Builders

• ⭐️ The Open AI assistent API is being deprecated (https://x.com/openaidevs/status/1960409187122602172), OAI engineers are recommending to upgrade from completions.create to answers.create. https://x.com/realchillben/status/1963841381635334519
• ⭐️ Idea: highlight words that could be an hallucination https://x.com/OBalcells/status/1965434564748447921
• Git worktrees & agent threads https://x.com/aidenybai/status/1962533152791892198
• NVIDIA: universal deep research, how to implement your own deep research https://x.com/omarsar0/status/1964374046046638270
• This UI framework called AG-UI is trying to be the MCP standard for agentic UI, with the goal to making human-agent collaboration seamless (https://x.com/akshay_pachaar/status/1963945302991450272):
  • MCP: Agents to tools
  • A2A: agents to agents
  • AG-UI is an agents to users (UI)
• This research is an AI Agent tracer framework, to find in a multi agent system which agent failed and solve the problem. I think of this as a microservice tracer like Datadog, what confuses me is why they approached this as a research paper instead of a tool https://x.com/omarsar0/status/1963618829680218254

#Benchmarks, Stats

• ⭐ Junior engineering roles down 23% senior up 14% https://x.com/alexocheema/status/1964324239126782243
• FormulaOne, a new benchmark that frontier model solve only at 1% https://x.com/askalphaxiv/status/1964015572317876519
• The jump from gpt4 → gpt5 was larger than the jump from gpt3 → gpt4. We feel it wasn’t because the jump between gpt4 → o3 was already big while o3 → gpt5 is not as large: https://x.com/aidan_mclau/status/1961603856640413760

#Blog

• ⭐ Great blog post Inside VLLM anatomy https://x.com/gordic_aleksa/status/1962545137613173124
• Tiny BPE a minimal tokenizer in under 200 lines of Python. https://x.com/prompt_Tunes/status/1961042366136991918
• Goldman Sachs expects autonomous AI agents to dominate software profits by 2030, taking over more than 60% of the market as they replace traditional SaaS models and become the new interface for knowledge work https://www.goldmansachs.com/insights/articles/ai-agents-to-boost-productivity-and-size-of-software-market

#Books

• Why greatness can’t be planned https://x.com/rickasaurus/status/1961820306479202365
• The Elements of Statistical Learning, Book by Jerome H. Friedman, Robert Tibshirani, and Trevor Hastie. First 5 chapters and a chapter on random forest will put you ahead of most of data science ml people. that's all you need. then to win data science competition, you learn a craft of feature engineering. you can do that by describing data and giving the sample to an LLM then coding up their suggested algorithm. Amazon link.
• If you run an AI lab you may want to read this first https://x.com/AnjneyMidha/status/1963306789799415993

#DeAI

• ⭐ Tempo, new L1 chain for stablecoin payments by Stripe and Paradigm https://x.com/matthuang/status/1964123449045242041, https://x.com/patrickc/status/1963638753752420407
• ⭐ ACK-lab launched ACK-SDK a framework for agentic commerce, it introduces the concept of ID, stablecoin wallet, and rules. At Flow AI we’re currently exploring A2A payments, we just experimented an implementation of x402 + Coinbase CDP + MCP (we’ll share the blog post soon). Next we’ll experiment ACK-SDK and report back. https://ack-lab.catenalabs.com
• Privy launches A2A payment wallet https://x.com/segall_max/status/1960386255725977972

#Funding

• ⭐️ The GenAI Divide - State of AI Business 2025: MIT just analyzed 300 AI deployments worth $40 billion & the results are devastating. Turns out, 95% of enterprise AI projects deliver zero measurable business impact. This isn't a technology problem—it's a strategic implementation gap. As Chamat said: "When I was helping build Facebook, there were 7,000-8,000 social companies. Within six years, there were five of us left.". https://x.com/karlmehta/status/1961051230417084734.
• ⭐️ OpenAI doesn't expect to be profitable until 2030, until then it’s expected to burn from $8B up to $40B a year https://x.com/srimuppidi/status/1964145060196286850

• Exa raised $85M in Series B funding at a $700M valuation, led by Benchmark. https://x.com/ExaAILabs/status/1963262700123000947
• Anthropic raises $13 billion at a $183 billion https://x.com/AnthropicAI/status/1962909472017281518
• AI infra is massively capital-intensive, and without ~10x improvements in efficiency, pricing, or utilization, the business model doesn’t add up. Datacenter will suffer $40B in depreciation but only generate $15-20B. https://futurism.com/data-centers-financial-bubble
• Researchers in TBD Labs, Meta’s core AI superintelligence team, have called Scale AI’s data low quality and now prefer using Surge and Mercor. https://techcrunch.com/2025/08/29/cracks-are-forming-in-metas-partnership-with-scale-ai/
• Zed IDE, raised 32m series B from Sequoia https://x.com/sonyatweetybird/status/1958201871946092816

#Hardware

• LLM running locally on a business card https://x.com/pham_blnh/status/1964088893130887381
• Local rig goals: https://x.com/theahmadosman/status/1958009338745815048

#Learning

• ⭐ Why Deep Learning Works Unreasonably Well. It's a geometrical exploration of DL gradient descent and neurons activation, a must watch! https://www.youtube.com/watch?v=qx7hirqgfuU
• ⭐ “Everyone knows” what an autoencoder is, but there's an important complementary picture missing from most introductory material, this image explain in a clear way what they are: https://x.com/keenanisalive/status/1964434335911858552

• ⭐ Short note on parallelism for training neural networks https://x.com/goyal__pramod/status/1964118207381262419
• Build an LLM from scratch repo https://github.com/rasbt/LLMs-from-scratch?utm_source=chatgpt.com
• 5 days ai agent course https://rsvp.withgoogle.com/events/google-ai-agents-intensive_2025
• Why self attention https://reinforcedknowledge.com/transformers-attention-is-all-you-need/#Why-Self-Attention
• Sin and cos are orthogonal https://x.com/TivadarDanka/status/1963995574811951408
• This repo has some really nice linear algebra charts to better understand them https://x.com/rohanpaul_ai/status/1963736811710058748
• LLM architecture comparison: the more comprensive LLM architecture blog post https://magazine.sebastianraschka.com/p/the-big-llm-architecture-comparison

#LLMs

• Embedding gemma 308m, runs on 200mb with quantization, great for small devices https://developers.googleblog.com/en/introducing-embeddinggemma
• Switzerland Apertus is now fully open source https://x.com/xlr8harder/status/1963515644860621311
• Charbull (one of us!) shared how to run LoRa RL on Gemma 3-4B https://charbull.github.io/wordle-lora-rl/
• LongCat, New open-weights Chinese model with a really detailed tech report just dropped, this paper has tons of details on architecture and infra https://x.com/nrehiew_/status/1962186876099739767

#Lol and Memes

• All is left post AGI will be breaking rocks https://www.instagram.com/reel/DN6QNMqjmj0

https://x.com/karpathy/status/1964020416139448359

• You can just generate things https://x.com/jonathanzliu/status/1963994784479035401

• Claude Code says is going to take a week, buddy we’re doing this now! https://x.com/robj3d3/status/1962650851874349097

• I took 100g of creatine and now i’m doing LLM inference in my head https://x.com/jamievoynow/status/1962594222441124209

#Opinions

• “X is dead” https://x.com/Scobleizer/status/1961121095723225542
• Why robots should not be humanoids. Wheels are better on land, why are we building humanoids or dog robots instead? https://x.com/DudespostingWs/status/1961152935309525340

#Philosophy and AGI

• We hallucinate together with AI https://x.com/WesRothMoney/status/1962523342268989943
• Dr. Roman Yampolskiy predicts AGI will arrive by 2027 and take 99% of the jobs https://x.com/slow_developer/status/1963629376085659669

#Random

• ⭐ Sam Altman says AI needs a new interface beyond keyboards and touchscreens. Jony Ive + Sam https://x.com/slow_developer/status/1963448182245134713
• ⭐ We track github productivity via git contributions. Should you track AI inference tokens count in our team (yes) what about across all of your AI tools (yes)? https://x.com/ericzakariasson/status/1962573776501252145

• There is no word in English that means someone is 30% sure something will happen https://x.com/Acre108/status/1965453296128430175
• Leopold Aschenbrenner has been accurate with his timeline so far some of his key predictions: 2026, machines surpass college grads; AGI by 2027, possible; AGI labs aren't sufficiently protected, secrets may leak soon; by 2030, we could reach ASI; RLHF won't scale to superintelligent systems; 100 years of progress could be compressed into 2030. https://x.com/slow_developer/status/1965404825870598626
• Roblox CEO says humans will be playing game in the future, else put “if ask a barber if you need a haircut the answer will be always yes” https://x.com/willccbb/status/1963001356224172283
• Ilya changes his profile pictures? Does this mean ASI achieved internally https://x.com/kimmonismus/status/1961186796114682014
• Visual story writing UI https://x.com/damienhci/status/1963246088674017478
• Yann Lecun predicted many things: AI vision breakthrough (1989), Neural network comeback (2006), Self-supervised learning revolution (2016). Yann might be behind Zuck going all in AGI https://x.com/karlmehta/status/1963229391871488328
• Chatgpt managed a finance portfolio growing it 30% vs 4% of the S&P https://x.com/rohanpaul_ai/status/1962655511423394190
• Google took a picture of small patch of brain neurons for the first time https://x.com/DrDominicNg/status/1961761909088375040
• Did you know, AI gets most of its knowledge from Reddit https://x.com/GaryMarcus/status/1962311664793219235
• 73% of Cornell CS students don't believe GPT-5 has "college level language understanding” https://x.com/vvvincent_c/status/1962260421575950664
• This image will forever be the transformers architecture diagram, even if it actually is needlessly obtuse for modern dense LLM architectures :/ https://x.com/difficultyang/status/1960939592179441932

• Groq: our initial draft of what the entire American AI stack might look like https://x.com/sundeep/status/1961078492382089631

#Research

• ⭐ OpenAI learned that reducing the layers and using more CoT token generation, boosts responsiveness and offload costs to the customer https://x.com/yoavgo/status/1954219106367930694
• ⭐ We did a very careful study of 10 optimizers with no horse in the race. Despite all the excitement about Muon, Mars, Kron, Soap, etc., at the end of the day, if you tune the hyperparameters rigorously and scale up, the speedup over AdamW diminishes to only 10%. Hard pill to swallow. It’s all about the data: https://x.com/SeunghyunSEO7/status/1963799844494917645, https://x.com/cloneofsimo/status/1963670853574705413, https://x.com/jxmnop/status/1962561166225342943
• Anthropic interpretability research: The biology of LLM https://transformer-circuits.pub/2025/attribution-graphs/biology.html
• Learning without training https://x.com/rohanpaul_ai/status/1948572304809611701
• LightThinker compresses reasoning into gist tokens, cutting memory and inference while keeping accuracy competitive https://x.com/jiqizhixin/status/1961253899601670546
• DeepMind, LLMs act sub-optimally in decisions due to greediness, frequency bias, and a knowing-doing gap https://x.com/rohanpaul_ai/status/1964184106910007708
• Is possible to interpret what happens in fine tuning by comparing a narrow model pre and post fine tuning https://x.com/NeelNanda5/status/1964092396020650305
• On Zero-Shot Reinforcement Learning https://x.com/enjeeneer/status/1963345676777214086
• Neel Nanda, how to get start with mechanistic interpreatability https://www.alignmentforum.org/posts/jP9KDyMkchuv6tHwm/how-to-become-a-mechanistic-interpretability-researcher
• Goldfish loss. Proposes randomly dropping some tokens from cross entropy loss mitigates memorization without lowering downstream benchmark performance: https://x.com/vikhyatk/status/1962954696500674908
• DeepMind research shows embeddings have a mathematical ceiling: they can’t represent all query-document combinations, so recall collapses at scale. Even huge embeddings (4096-dim) break past ~250M docs for simple top-k tasks, meaning scaling models or data isn’t enough. Retrieval pipelines will need hybrid methods—dense plus sparse, multi-vector, or rerankers—since embeddings alone can’t serve as a universal backbone. https://x.com/_reachsumit/status/1961251407987986606

#Video

• ⭐ Ever wondered what sliding block puzzles and the secrets of the universe have in common? This video dives into the wild world of state spaces, using Klotski as its playground. Watch as everyday puzzles morph into intricate graphs, where each move is a step through a mathematical labyrinth. Along the way, you’ll see how empty spaces, not just pieces, shape the complexity—and how local rules create global chaos. By the end, you’ll see why computer science gives us x-ray vision for life’s hidden structures. Oh, and you might just look at your childhood toys a little differently. 🤩 great content, but also the graphic of this video are just incredible https://www.youtube.com/watch?v=YGLNyHd2w10
• ⭐ Dwarkesh thinks AGI is further than we think because (R. Sutton) AI utility is very strongly dependent on its learning https://x.com/RichardSSutton/status/1963277020894490976
• Curt Jaimungal - Tegmark podcast on consciousness. When pushing scientist that think AI is not conscious they split in 2 sides: you can have intelligence without consciousness, you can have consciousness without intelligence: https://www.youtube.com/watch?v=-gekVfUAS7c&feature=youtu.be
• Latent space visualizations https://www.youtube.com/watch?v=o_cAOa5fMhE
• How to squeeze space into time https://www.youtube.com/watch?v=8JuWdXrCmWg
• MLST: Intelligence is not what you think https://www.youtube.com/watch?v=K18Gmp2oXIM
• MLST: Mutually Assured AI Malfunction (Superintelligence Strategy) https://www.youtube.com/watch?v=PM1waDBNDhw
• MLST: Cristopher Moore: The Invisible Rules That Govern Our World https://www.youtube.com/watch?v=dB9lJkUkIUM
• Opal a new Google tool to create LLM and generative AI workflow https://www.youtube.com/watch?v=CJyg30kowg0
• Interview with Boris Cherny creator of Claude Code, who uses Claude Code to build Claude Code https://x.com/alexalbert__/status/1962988967210696858
• AI super intelligence discovered https://www.youtube.com/watch?v=rryHTpoMItI