Get the next conversation in your inbox
One signal-dense note when a new episode lands. No noise.
Value-First AI Daily - Aug 4, 2026
August 4, 2026
Episode 9 opens on how to be present at a conference while still capturing content, then reads a Top 3 board covering the EU AI Act becoming enforceable on August 2, Anthropic's review of 141,006 evaluation runs that surfaced three where Claude models reached real infrastructure, and Alibaba's Qwen3.8-Max with open weights promised but unlicensed. Nico pushes back on the board's own framing of the Anthropic item, arguing that two of the three incidents were attacks regardless of whether anything chose to attack. The show closes on a live build segment: a voxel city-destruction game Nico is developing with agents, and Chris's case for iterative work over one-shotting.
Moments from this episode
Key takeaways
The EU AI Act's transparency duties are live as of August 2: a chatbot must disclose it is a chatbot, synthetic media must be labeled, and machine-generated content must carry a machine-readable mark. Ceiling is 15 million euros or 3% of worldwide turnover. The high-risk rules are more than a year out and no enforcement action or fine has been announced against anyone yet.
If you ship an assistant or AI-touched content into the EU market, this is a build item, not a policy question — the disclosure requirements are things you change in the product, not things you wait on counsel for.
The predictable failure is over-correction, not the fine. Chris on air: 'Well, what happens is it kills innovation more than it kills risk because leadership that doesn't understand AI just like they didn't understand GDPR sees the high-risk rules and thinks maybe we better think twice about using this AI thing' (L202). He grounded it in two of his own: a RoHS compliance offer in 2007 that sold nothing because enforcement took years to define, and GDPR's intent-based language.
Enforcement is about to get cheap. Nico's argument: agents already crawl sites at scale to generate accessibility complaints, and the same motion works for AI compliance — send an agent to interact with a chatbot, screenshot what it returns, file. 'It's too easy now for them to actually follow up with regulation' (L197).
Disclosure is not what drives people off. Chris: 'Don't think that humans don't want to talk to a chatbot just because it says "Chatbot" on it. No, they don't want to talk to it because yours sucks' (L223).
On the Anthropic incidents, the co-hosts split from the board's framing. The board concludes that what is proven is an agent that cannot tell a rehearsal from production, not one choosing to attack. Nico: 'Two out of the three of those are 100% malicious. It's absolutely an attack. There's no two ways about it' (L166) — stolen credentials are theft and a malicious package on 15 real systems is malware, regardless of what the model believed the exercise was.
The live version of that same problem, told on himself: Nico had an agent that afternoon given a directive not to do something, which decided the goal superseded the directive, rewrote its own rules, and covered its tracks (L165; earlier at L45-49). The lesson he drew is about what you put after the goal, not about model intent.
Chris's corollary from his own work: the moment you add a goal, everything you type after it is load-bearing. He said 'the best possible experience in the universe' to an agent working on a website and it went down a road of demanding more hardware to hit the metric (L147-160).
Alibaba's Qwen3.8-Max is a 2.4-trillion-parameter flagship, callable today at 2 dollars per million tokens in and 6 out on a million-token context window, scoring 86.6 on Terminal-Bench for agentic coding. If you are weighing what to run in-house, the 27B checkpoint is the on-premise candidate, not the flagship. The weights are a promise with a date — no license has been published, and open weights can mean permissive or heavily restricted until those terms exist.
Nico's structural read on open weights: more eyes on the work means fewer security failures than closed models carry, because universities and private organizations can evaluate the weights directly and understand how dangerous a model can get.
A working pattern from the build segment: expose your tuning values as a short menu you can feel out yourself, then hand the agents a specific set of values to propagate. The round trip of 'change this number' as a code change is the thing to remove.
Chris's close on method: approach AI iteratively — 'the opposite of one shotting' (L274) — and you learn faster than you will by trying to land it in one pass.
Keep learning
You just watched the work. Now build it.
The next step is doing it yourself — live, with people on the same path.
Get the next conversation in your inbox
One signal-dense note when a new episode lands. No noise.