Here are this weeks links. The weak was headlined by the release of GPT-5 but there was a lot more happening besides that:
- GPT Launches, Model Card, and Overview. GPT finally launches this week to much fanfare, some calls of disappointment, and new features. The consensus is that the model improves things, but is not wildly better. OpenAI has done a number of interesting things, though: GPT-5 now decides internally which model to use to respond to your query, and consumers, now no longer chose which model to query (this has caused some concern over predictability, though older models remain available to developers). There will probably be growing pains with this, but it is a logical step for OpenAI and evidence of its consumer focus. Asking users to choose which model to use is simply counterintuitive, and by making the decision internally, OpenAI can manage costs more efficiently. Another interesting feature is the ability to select the "personality" of responses: from robotic (terse) to nerdy (curious), to cynic. I already feel like I'll need a "mix" of these depending on the query type. You can learn how to select personalities here. There is also a new prompting guide.
- McKinsey and its peers need a new strategy. And some humility. A short piece in The Economist about one of the key interesting questions in AI and Business. How will AI affect consultancies. The arrival of new technology has been a massive boon since all of a sudden every major company is scrambling to rethinkg strategy and best practice. Consultancies are well placed to try to bring together ideas from all spheres and help implement change. In the long run however, will AI be a better consultant. A system embedded in the firm with constant access to data, able to draw on global knowledge and continually being able to optimize processes over time, probably beats external consultants at many tasks.
- Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens. An interesting analysis of Chain-of-Thought techniques in LLMs. The brutal conclusion is that they are brittle and near useless when the model is perating on questions outside it's trainign distribution. This is not really suprising since Chain of Thought really isn't using a model of the world to understand a problem, it's just searching for threads of connections similar to what was asked. This doesn't mean Chains-of-Thought are totally useless. However it does mean that the "illusion" of reasoning they can provide may be quite misleading.
- “Supersonic Growth”: Base44 set to hit $50M ARR just months after Wix acquisition. Breakthrough new AI features + distribution is a potential pattern for sucess, if done right. Wix aquired Base44 a few months ago, with Wix management no doubt seeing the threat to its business from Vibe coding platforms such as Lovable and Bolt. Now Base44 is already adding a %age point of two to Wix' overall 2025 growth projections. No doubt this due to the multiplier effect of new tech that already works well with a huge distribution advantage that Wix has across it's user base. Watch for the newly public Figma to make some interesting AI acquisitions well...
- The Real Reason 50%+ of Our Sales Team is Now AI (And What We Learned). I think this may be one of the most important AI economy posts of the last few months. Jason Lemkin and the team at SaaStr have constentily tried to push their own AI adoption and this piece talks about why they are making certain choices. The key important take-away is partly chilling but also clarifying: for many roles it's relentless execution and organization that matters. In those roles, unfortunately humans actually don't do well and managing the humans involved creates huge overhead. The AI replacement don't actually need to be better... they just need to produce less drama.
Lastly ... OpenAPI also released Open Source models! More on that in a future post!
In other news the robot launder Perseverance took a super cool vista of the Martian surface.
