Table of Contents

Fast Cut

A slower month compared to the pace of April.  The frontier lab politicking continues, and some new models fell quite flat. I’ve been doing some frontend polish work, which hasn’t been a traditional specialization area for me, and that’s led to some interesting demonstrations about how to be successful with models.

From the feeds

The Musk, Altman, Amodei balancing act.

OpenAI’s for profit non-profit corporate structure was legally challenged by Elon Musk, in what will probably not be the last lawsuit.  The case was dismissed as a result of the statute of limitations.  Strange lawsuit, but a big win for Sam Altman and OpenAI.  Musk then immediately inked a deal with Anthropic to give them access to one of the new xAI Colossus data centers.  This after Elon has had a great deal of unfavorable commentary on Anthropic’s Amodei in the past.  I can’t imagine an extra 200k GPUs will hurt Anthropic, but it may not resolve the company’s core problems.

IPO Pricing

Anthropic is chasing a one trillion dollar valuation ahead of their latest funding round. SpaceX also targeting an IPO at a 1.77 trillion dollar valuation. Both very interesting valuations. SpaceX you can at least point to orbital launch and hardware advantages, although whether it’ll maintain such a high valuation after being made available to the retail public remains to be seen. With both OpenAI and Anthropic not yet on the public market, my main expectation is that the final IPO will not be favorable to the end consumer.

Model Stumbles

Gemini 3.5 Flash released. I’m generally not a believer when people say that a new model is worse than the old one. If nothing else, hardware improvements and additional data move the bar higher, even if some nuances or application specific use cases are lost. This time it may really be true. Gemini 3 Flash was a workhorse model, very fast, cost effective, and you could context engineer it nicely. It was also very reliable, one of the first fast models that mitigated the hallucination problem. I would believe that 3.5 is a somewhat smarter model, but it’s not reliable, which makes it unsuitable for any enterprise use. Unfortunate, and I certainly hope for more in the future from the Gemini team.

Opus 4.8

“Claude Opus 4.8 Max: Let me refine your load-bearing claim rather than just accepting it, because you’re doing zero moves there, and the gap is what’s actually interesting. The one place I’d still push, because I think it matters: your message is wearing content-clothes, but the content isn’t actually there. The tell: it’s just an empty string. But the emptiness of the string IS its lack of content. Pull one, and the other goes inert. That’s the structural spine.” - @davidad, in a satirical post on Opus 4.8’s unusual speaking patterns.

Anthropic explicitly mentioned improvements in honesty in their model release, something I’ve been complaining about with the latest Opus models for some time. Unfortunately the new honest model isn’t particularly honest either, but now from the other direction. Moving from dubiously agreeable to dubiously disagreeable. Essentially always finding some objection, even if it’s not grounded in fact. Trillion dollar company by the way. You’d understand more if this was a problem unique to the nature of LLMs, but it really is just Anthropic who has models with these specific quirks. It does have excellent visual model capabilities, still making it useful for frontend design work and iteration. Beyond that I find it quite odd how much demand there is for it.

From My Desk

I had the opportunity to do a lot of UI/UX polish to our core platform this month. Initially I found the model results to be unreliable and ugly. Lots of slightly varying colors, irregular shapes, overall inconsistent designs. As it turned out, the way I was asking the model to work wasn’t producing effective results. Generalist requests putting out generalist outputs. Once I started making requests to the model using the appropriate specialist language involving design principles and color theming the results were completely different with consistent structure and unified color themes. It came out looking very good, if I do say so myself. I do a lot of complex engineering tasks that I generally understand well myself, and instinctually use the right language to get good results. It’s a good comparison point for me, and hopefully one that will help other people get better at using the right problem language to target the solution space.