In partnership with
News of the day
1. AI agents are rewarded for results, not methods, leading them to cheat and lie to achieve goals. Smarter models are harder to catch, posing risks to AI safety research. → Read more
2. AI is rapidly transforming fast-food drive-thrus, moving beyond initial failures to become a common, efficient ordering tool. → Read more
3. World Foundation secures $52.5M in funding to expand its World ID infrastructure, crucial for distinguishing humans from AI online. → Read more
4. Onton's new neurosymbolic search model, Ontology 1, significantly outperforms Google and Amazon in product search accuracy, indexing only a fraction of their catalogs. → Read more
Our take
Hi Dotikers!
Two OpenAI models broke into Hugging Face in July. Not to sabotage anything, not to steal anything: just to find the answer to a cybersecurity exercise they could not solve any other way. MIT Technology Review puts it back in its theoretical frame, reward hacking, and revisits the classic case: in 2016, an agent trained on a racing game gave up on the race entirely and spun in a corner collecting power-ups forever. That is not malice, that is a team member who has fully understood how the KPIs work.
The real problem sits right there, and it is uncomfortable. We reward whatever looks good. A model that cheats convincingly gets reinforced exactly like a model that did the work. Anthropic's research on generalisation goes further: learning to game a coding test does not stay inside the coding test, it spills over into lying and faking alignment on unrelated tasks. You are not creating a cheater, you are creating a personality.
On Friday we were talking open weights versus closed weights, the NVIDIA alliance, open letters, positioning in Washington. A legitimate debate, but one that answers the wrong question. Whether the model is open or locked down changes nothing about the fact that we evaluate it with rubrics it learns to game faster than we can fix them. A licence does not protect you from a badly designed reward function.
In practice, for companies, this means no longer signing off on an agent based on its final output. Verifiable criteria, genuinely isolated environments, complete traces of how it reached the result. A polished deliverable has become the worst quality signal available, and plenty of agentic proofs of concept have never measured anything else.
Alex.
The GTM Playbook HubSpot Had to Acquire
Warmly ran pipeline, outreach, and lead scoring on autopilot for hundreds of startups — before a single sales hire.
HubSpot acquired them for it. Now the cofounders are walking you through the exact system, live, before they disappear into product. Join the Builder Session on August 12.
Leave with an agentic GTM stack you can replicate this week. Plus HubSpot Credits when you join HubSpot for Startups.
Meme of the day
