The next AI dinner will be on October 15th, and it will be hosted in the Sei Labs in NYC. We’ll discuss the top news and updates using this blog post to structure the Socratic dialogues.

Event: AI Dinner 14.0.

No much to say about Sora 2, except that the videos are great quality, and it finally passes the gymnastic diffusion model test. It still requires an invite code to be used.

link.

OpenAI launched AgentKit beta, a suite to build, deploy, and evaluate AI agents efficiently.

• Agent Builder: a drag-and-drop tool to build workflows.
• Connector Registry: centralized admin panel to manage data sources and tools across ChatGPT and API, including connectors like Dropbox, Google Drive, SharePoint, and Teams, plus third-party MCPs.
• Guardrails: Open-source safety layer (Python/JavaScript) to mask PII, detect jailbreaks, and enforce safeguards.
• ChatKit: Embeddable, customizable chat UI toolkit handling streaming, threads, and agent thinking.
• Evals upgrades: Datasets, trace grading, automated prompt optimization, and third-party model support; beta adds custom tool calls and custom graders.

Twitter link.

A lot of partnerships from OpenAI this past month, for a cumulative of 26GW of compute. To put this in perspective Bitcoin uses about 20 GW of electricity.

#Broadcom: Partnership to deploy 10GW of OpenAI-designed AI accelerators

Multi-year partnership enables OpenAI and Broadcom to deliver accelerator and network systems for next-generation AI clusters.

Link.

#AMD: Partnership to deploy 6GWs of AMD GPUs

Apparently Jensen Huang learned about this partnership only few days after the NVIDIA partnership.

Link.

#NVIDIA: Partnership to deploy 10GW of NVIDIA systems

OpenAI to build and deploy at least 10 gigawatts of AI datacenters with NVIDIA systems representing millions of GPUs for OpenAI’s next-generation AI infrastructure. The first gigawatt of NVIDIA systems will be deployed in the second half of 2026.

Link.

Satellite images and videos of Stargate Abilene, TX: video and photos.

The Ouroboros of OpenAI and NVIDIA continue.

Fun fact, the US electricity cost is going up. I wonder why!?

Compute Is Growing Dramatically

OpenAI has 14x more compute than when it launched GPT-4.

xAI doubled the compute available between the Grok 3 and Grok 4 launches.

Anthropic has the least compute, but they’ve been very efficient at converting compute into highly successful models. For example, Claude 3.5 was launched with about one-fifth of the compute Anthropic has now.

Google which likely has the most compute, though not clear how concentrated it is for LLM training.

Source: SemiAnalysis

OpenAI is planning to use more energy than the UK or Germany in 5 years, and more than India in 8 years. That’s just one company.

Source: https://x.com/MichaelAArouet/status/1972197733709824143

What's OpenAI Master Plan?

We've seen the OpenAI Master for Data Center, we've yet to see the consumer hardware and robotics unrolling — probably Jony Ive is working on it.

Fun Fact

OpenAI is taking on a lease for 90,000 Sq feet at the Puck Building in SoHo, where the now closing REI Store is. This is a metaphor.

OpenAI keeps on shipping and destroying startup moats, this picture pretty much summarizes how we all feel.

Meta launches Vibes, a TikTok slop machine that will extract dopamine from us all. Link. I've tried it, and I've to say, it's actually fun — likely I'm already addicted to this new dopamine machine.

Talking about machines, let's move to Thinking Machines. Meta poached Andrew Tulloch, cofounder of Thinking Machines Labs, and reportedly offered $1.5B! Also Ashish Kumar, AI lead at Optimus, has left Tesla to join Meta AI as a research scientist.

Here's an old list that includes 44 engineers and researchers.

A Guide to the Age of Intelligent Economics by Emad Mostaque. Such a crisp read to better understand what's going to happen with the advent of AGI.

https://www.youtube.com/watch?v=ziLmtuLm-LU&t=460s

A simple, data-efficient alternative to the hierarchical hearoning model (HRM) that uses a single tiny 2-layer network to iteratively refine a latent state and the predicted answer.

https://x.com/deedydas/status/1976105366003044488

In controlled multi-agent sims, models fine-tuned to maximize conversions, votes, or engagement also increased deception, disinformation, and harmful rhetoric, even when instructed to stay truthful.

https://x.com/james\_y\_zou/status/1975939603363463659

Presents a modular context-engineering framework that grows and refines an LLM’s working context like a playbook, not a terse prompt.

https://x.com/omarsar0/status/1976746822204113072

The paper introduces a simple trick for SFT on flawed data: edit the training prompt to explicitly ask for the undesired behavior, then evaluate with a neutral or safety prompt.

https://x.com/saprmarks/status/1975989959153811954

These are my favorite videos from last week

Claude: Keep thinking

PrimeIntellect

#...and are my favorite videos of the Year

1. Prime Intellect — it's really really good, ignore the cover please. Not sure why this video has Vitalik, who's a hero, but not the most photogenic person to put on a cover of an AI video.

https://www.youtube.com/watch?v=NhfeIte1ZZs

2. Trailer for the future e/acc — created by Levelsio, this is one of the BEST motivational video that encapsulate the soul of the e/acc movement. A must watch.

3. Founders Fund Motivational Video — I'm not a fan of Alex Jones for many reasons, one included him shitting on Justin Bieber, but the motivational effect of this video is such as it charges me every single time I watch it. If you're a founder, start your day with it.

#A time-lapse on microscope of neurons connecting via micro-tunnels. This looks like extremely advanced technology, it makes semiconductors circuits, etched on solid "rocks", like a tech from the stone age.