Michael Buckbee

The real value in AI is the non state of the art models. Here's the same task with the new Gemini 3.5 vs some cheaper GPTs. Very similar quality, 75% faster and 98% less expensive!

Here's the quick demo if you haven't seen it.

For HutchDB.com I'm still struggling to come up with a tagline. Help.

@Timur_Yessenov Yeah, I do a lot more of this with Claude code. And to me the important thing is structuring you know with lots of gates so that even if something does get messed up it's still confined to development and blocked by lots of tests and merge approvals.

@mbuckbee The mobile piece makes approvals feel lightweight, but it needs a custody screen: repo, machine, current diff, risky command pending, and timeout. Otherwise “approve from phone” becomes approve blind.

Codex on your phone: search becomes doing OpenAI shipped Codex inside the ChatGPT mobile app this week. You can monitor and approve coding tasks from your phone while the agent works elsewhere. Developer story on the surface, broader shift underneath.

Source: openai.com/index/work-wit… Full issue: knowatoa.com/disrupted

If your marketing depends on being found during the lookup step, the only way to stay visible is to be part of what the agent already knows. Back to the training-data and reachable-page problem.

Most AI hallucinations aren't the AI's just wilding out, it's a "gap" a space where they didn't have _any_ info to go on so they filled in the lines. If you're running a business you should proactively be filling those in by publishing content that directly states what it is you do, what problems you solve and who you help.

Source: openai.com/signals/resear… Full issue: knowatoa.com/disrupted

This is a trend we're seeing across all the different AI tools, not just ChatGPT, where AI-assisted search and decision making is becoming mainstream.

ChatGPT's audience just aged ten years I'm always interested in the demographic studies of who's using AI tools. They tend to surface unexpected uses and people jumping in for new and interesting problems I didn't even know existed.

@LeahHun46167833 Yeah, we (Knowatoa) do that as well.

ChatGPT just cut 20% of the sites it cites New research on how ChatGPT search works found that after the GPT-5.3 switch, the model cites about 20% fewer websites per response. The citations that survive go to higher-authority domains.

Source: lilyraynyc.substack.com/p/it-works-unt… Full issue: knowatoa.com/disrupted

The version of AI content that works is the one where the human brings the ceiling and AI helps with everything below. Not coincidentally, this is exactly the approach we take with Knowatoa and our research-heavy content generation.

Data on how to win with AI content Lily Ray published research this week looking at what works (and what doesn't) when you scale AI content. Go read it. The hot takes I've seen on socials about the article miss a lot of the real thought and nuance that went into the analysis.