Parameter Update: 2025-40

Slowest week in a long while, but Veo 3.1 is still very cool!

Veo 3.1

While Sora is still covering my timeline in (sometimes surprisingly good) AI slop, Google has stepped up their game this week with Veo 3.1, adding better scene control (being able to provide both start and end frame is not technically new for video models, but is sick) while further improving video quality. In my test, the result seems just a bit better in every dimension. The real story, for Europeans at least, is access - while I still haven't got to try Sora, Veo was just immediately available through the Gemini web app and Flow (which works surprisingly well!). I just wish the rate limits were a bit higher.

Below, I've given it these two reference images

and the prompt "The man walks into a coffee shop, orders and sits down". Here's the result:

Overall, I applaud Google for resisting the urge to immediately release their Vertical Video Slop App, though some of the stuff going on with YouTube Shorts has me worried.

Anthropic

Claude Haiku 4.5

After releasing Sonnet 4.5 a few weeks ago (in my experience, the most pleasant LLM to talk to yet) this week Anthropic announced it's smaller brother, Haiku 4.5.

This one apparently matches Sonnet 4 in SWE-bench, which according to Anthropic means it "matches its coding performance at one-third the cost and more than twice the speed". While I haven't tested it myself (I haven't used Claude Code ever since they butchered rate limits a while ago and probably won't switch back until OpenAI takes away my Codex rate limits), it's probably more of an indictment of SWE-bench, given that like half of it is just django?

I'd also be remiss to note that they're pricing it at roughly 4x what Haiku 3 used to cost 1.5 years ago. I thought these things were supposed to get cheaper over time?

Skills

While OpenAI has just extended MCP with their "Apps in ChatGPT" feature, Anthropic seems to be building into a slightly different direction. "Skills" are effectively just a bunch of folders containing SKILL.md markdown files (and, optionally, other materials required to use them).

On the one hand, this feels surprisingly simple, and I am not entirely sure how it fits in with their own existing MCP ecosystem, on the other hand I had similar thoughts about MCP back when it was first introduced, and while that specification is far from perfect (and was massively overhyped for a bit there), it did turn out to be tremendously useful?

DeepSeek OCR

This was just revealed earlier today, so I haven't managed to take too deep of a look at it yet, but it seems that DeepSeek may have just pushed state-of-the-art OCR forward significantly! Apparently, the compression ratio on their vision token is much, much higher than for previous models without loosing accuracy, making it nice and fast. Either way, OCR feels like it's been massively underdiscussed, so I'm excited to take a proper look at the (usually excellent) report.

OpenAI

Broadcom Deal

After announcing plans to build out "at least" 10 gigawatts of Nvidia GPUs and 6 gigawatts of AMD GPUs, OpenAI has now announced plans to also build out 10 gigawatts of custom accelerator hardware in collaboration with Broadcom. For context, this comes at a time when

• Meta is announcing 1MW projects,
• McKinsey projects 170–220 GW total, worldwide datacenter demand by 2030, and
• Microsoft is starting to cool down to the idea of building out more datacenters (but also weeks after construction began on a nuclear fusion plant to power their new infrastructure by 2028).

I know calling the AI infra build-out a bubble is overdone at this point, but they're really not making it easy.