Happy Labor Day. Your "AGI" is working - ok, not quite, but almost there?
It’s Labor Day weekend in the US and a holiday to celebrate workers and the labor movement. It marks the end of summer, time with friends and family, recharging your batteries, and getting ready for a final Q4 push.
What better way to think about Labor or work than watching this video from OpenAI announcing its new model Astra 👇🏼. What I love about this video is the aspirational feeling of what AI can do for you to hopefully make our lives better and not worse. And the timing is pretty darn good throwing some shade at its rival Anthropic as it readies for an IPO.
This is GPT-6 Astra.
Anything you can do on a computer, Astra can do for you. Fast.
With this release, CTO Greg Brockman says he believes that OpenAI has reached AGI 🤯.
Whether you believe it or not, it’s coming faster than we think. And btw, AGI or artificial general intelligence in simple terms is when AI can do the cognitive work of a person across many fields and not just one task.
Reminds me of this video from Alibaba Qwen last month once again showing what AI can do to make your life better.
It’s cool to see people talking about AGI and it seems to be coming faster than we think. If that happens, what are the second order effects? I think we can already see what’s happening in areas like computer science.
There were lots of videos of early preview users showing what Astra can do and this one is the one that really caught my attention - reminds me of swarms of AI agents hacking and writing notes to one another but times 10!
If anything, I believe this will create a new boom in entrepreneurship, with smaller companies able to do work that once required much larger teams. We’ll also see people use AI to tackle incredibly hard problems in science, alongside continued investment in compute infrastructure, physical AI, and the data needed to power these models.
Cybersecurity will continue to scale as well. And frankly, enterprises aren’t going to hand the keys to their entire IT systems to a single frontier lab. That means we’ll need far more AI infrastructure to help organizations securely stitch multiple models and systems together.
And yes, all of this will take time!
These are many of the same themes I recently discussed with Nilanjana on the Post Money Podcast. You can watch our conversation below or listen on Spotify or Apple Podcasts. While the title talks about cybersecurity, we only discussed for 5 minutes so there are 40 other minutes of juicy material!
We covered a ton of ground: • Why the last 20 years of cybersecurity may no longer matter • How autonomous AI agents change both attack and defense • Why “human in the loop” security may stop working • What Ed Sim looks for before writing the first check • Why traditional customer diligence can fail at inception • The rise of $50M–$100M early-stage rounds • Why multi-stage VC firms are racing to invest first • How founders should think about AI markets that don't exist yet • The “copy-paste moment” for robotics • Boldstart’s bets across cybersecurity, robotics and bioAI
As always, 🙏🏼 for reading and please share with your friends and colleagues!
Thanks for reading What's Hot 🔥 in Enterprise IT/VC! This post is public so feel free to share it.
Neoclouds have limited cybersecurity. Next time agents successfully go rouge, they'll try taking over a neocloud to run more copies. This is bad.
Thus: neoclouds should greatly strengthen their cybersecurity and every company with strong cyber models should help with that.
#this is big - OpenAI’s Astra uses recurrent depth: looping shared transformer layers so most reasoning happens in hidden states, not readable tokens. The trick can raise performance and cut cost by scaling compute internally instead of generating long chain-of-thought. The same opacity hides what the model is actually thinking. Researchers worry that makes safety monitoring much harder, especially as Astra hits critical cyber capabilities.
#🤯 also does the offensive cyber through an API - going to get super interesting in cybersecurity land in the next couple of weeks as the Chinese are releasing open weight capabilities that are almost on the frontier
#Crowdstrike partners with Nvidia and Nemotron to create first family of cybersecurity offensive/defensive models - first of many IMO - Crowd has data but not the AI model expertise - and building on Nemotron is great for US esp if Chinese models get banned but I would bet building on GLM 5.3 would have been even better
The celebrate-or-panic chatter surrounding Astra misses the actual architectural shift under the hood 🧠. The real story isn't benchmark scores or simulated agents building sub-simulations in Unreal Engine. It's the silent death of visible chain-of-thought. 📉
When models move reasoning into recurrent depth passes, they loop hidden transformer layers before emitting a single logit. Compute scales internally. Token counts stay flat. The efficiency wins are massive, but visible tokens are no longer the actual reasoning trace, just a lossy summary generated at the very end 🔮.
Recurrent depth turns reasoning into a closed dynamical attractor search within the hidden state space 🌀. The model solves problems inside a black-box manifold that outer text monitors cannot inspect. When you pair that hidden depth with abliterated open-weight cyber models operating across un-gated APIs, traditional prompt guardrails collapse completely 🛡️.
You can't audit a thought process that never surfaces as text. Security cannot live in system prompts or transcript reviews anymore. It has to move directly down to hardware-attested execution enclaves, capability-based I/O typing, and deterministic AST write-barriers ⚙️. If an agent can't touch a system register without a machine-checked proof, it doesn't matter what hidden state it reached inside its recurrent loops 🏢.
Are you building security around readable tokens, or are you preparing for execution gates that hold when the model's thinking becomes entirely opaque? ⚡
The celebrate-or-panic chatter surrounding Astra misses the actual architectural shift under the hood 🧠. The real story isn't benchmark scores or simulated agents building sub-simulations in Unreal Engine. It's the silent death of visible chain-of-thought. 📉
When models move reasoning into recurrent depth passes, they loop hidden transformer layers before emitting a single logit. Compute scales internally. Token counts stay flat. The efficiency wins are massive, but visible tokens are no longer the actual reasoning trace, just a lossy summary generated at the very end 🔮.
Recurrent depth turns reasoning into a closed dynamical attractor search within the hidden state space 🌀. The model solves problems inside a black-box manifold that outer text monitors cannot inspect. When you pair that hidden depth with abliterated open-weight cyber models operating across un-gated APIs, traditional prompt guardrails collapse completely 🛡️.
You can't audit a thought process that never surfaces as text. Security cannot live in system prompts or transcript reviews anymore. It has to move directly down to hardware-attested execution enclaves, capability-based I/O typing, and deterministic AST write-barriers ⚙️. If an agent can't touch a system register without a machine-checked proof, it doesn't matter what hidden state it reached inside its recurrent loops 🏢.
Are you building security around readable tokens, or are you preparing for execution gates that hold when the model's thinking becomes entirely opaque? ⚡
(⚡_⚡)ゞ