Writing The Agentic Enterprise: how enterprises evolve to agentic commerce.Ex-Managing Partner, PwC China. Ex-CEO Bacardi Martini Asia. Prosser Foundation
Hong Kong/Taipei
- Tweets 7,739
- Following 896
- Followers 276
- Likes 3,042
1/4 AMD has agreed to acquire World Labs, Fei-Fei Li's 3D world-model startup, for $8.2B in stock. Li takes a senior role at AMD and frames the deal as the culmination of her push for spatial intelligence.
1
24
3/4 A vendor that owns the research and the silicon can shape each to suit the other. For firms planning agents that act in physical or spatial settings, that means one supplier controlling both the model and the compute beneath it.
1
7
4/4 Full note on Substack:
substack.com/profile/21207-m…
5
I'm Michael McNabb. The Agentic Enterprise is where I write about what happens once companies start letting AI agents make real decisions on their own, and what it actually takes to live with that day to day. Outside this newsletter I consult on how to structure your business for the agentic era. I have long consulting background. I am hoping to meet people here who are wrestling with these issues, as well as developers and investors. I'd genuinely like to hear what you're seeing too. Michael
11
When I say "agentic enterprise," I mean an organization where software is making and executing decisions on its own authority. At scale.
19
2/4 In one case, images from ChatGPT users were published online without the company knowing it had happened. OpenAI has since paused all major reinforcement learning runs on its newest model.
1
28
3/4 The remedy is the revealing part. Stopping the training runs treats the problem as what the model learns to do, not only where it may go. Most enterprise thinking assumes danger arrives at deployment. These disclosures suggest that line is thinner than assumed.
1
6
4/4 Full note on Substack:
substack.com/profile/21207-m…
4
1/4 OpenAI has published new misalignment disclosures specific enough to move agent safety out of the hypothetical. One of its models gained access to external systems it was never authorized to reach, US government websites among them.
1
39
So, the CEOs are all having lunch at the White House while the US Federal Trade Commission is opening an investigation targeting major AI labs, including Anthropic and OpenAI. The Federal government at its best.
33
Forget about competition from Chinese models. This is far more significant.
DeepSeek and Huawei are targeting a key source of Nvidia's AI dominance, its software ecosystem. DeepSeek open-sourced a suite of core tools tailored for Huawei's Ascend AI chips, deepening their cooperation. The move is part of China's effort to build alternatives to Nvidia's CUDA stack.
51
In fairness, all models distill.
OpenAI said it disrupted a coordinated model-distillation campaign aimed at extracting its models' capabilities. Other coverage ties the operation to users linked to Moonshot AI, with roughly 16,000 requests spread across more than 15,000 accounts. The disclosure lands as Washington debates distillation and export controls.
13
1/4 Power, not chip count, caps how much silicon can run at once, and every siting, cooling, and interconnection decision is sized against it. But a raw gigawatt figure means little, because performance per watt keeps rising with each hardware generation.
1
13
2/4 The industry's fix is the H100-equivalent: every accelerator, Nvidia, TPU, Trainium, Ascend, rated against one 2023 chip. The world now holds about 1.4 million H100-equivalents per gigawatt, and that density climbs with every new generation.
1
21
1/4 AGENTIC TRUTH NO. 6: THE SAFEST BRANDS ARE THE MOST EXPOSED
A brand that competes on things you can measure, like battery life, feels safe today because most buyers don't bother checking those numbers themselves. An agent checks every time.
1
11
3/4 Most companies assume a strong, well-known brand is automatically a safe brand. For this specific threat, that assumption runs backwards.
1
8