2026-03-05

Trista and I saw NIN last night with friends, and it was a great show, as always.

I inevitably wind up thinking about the gap between the live event experience and reproductions on video or in VR. The gap is substantial, with the “shock and awe” of high powered, extreme dynamic range stadium effects, but there are still levers to pull.

A capability of the Quest VR hardware that would be fun to exploit is per-frame varying of the display persistence. The low-persistence display is only flashed “on” for about a tenth of each frame, so it doesn’t look blurred as you move your head around. This is a tradeoff between screen brightness and blur, which is just statically picked by Meta today, but it could be done dynamically on a per-frame basis (coordination would be tricky!).

Going to full persistence briefly would give an instant 10x brightness increase, which would be quite impactful on eyes adapted to the 10% brightness. You wouldn’t want to keep it there for long, because the blurring would be obnoxious and eyes would adapt to it, but it would be perfect for the epilepsy-inducing concert light flashes.

The clean way to do it would be with a 16 bit high dynamic range linear buffer that gets factored into an sRGB image for the display and a backlight flight time, but a hacky extra parameter latched with the frame submit would be the easiest thing to experiment with. Make it work, then make it clean!

A tech demo would be fun, but Beat Saber has a large enough user base that stepping up “concert lighting effects” might even be worth it from a net user value standpoint.

Trista: NIN - Friggin 🔥

Trista and I saw NIN last night with friends, and it was a great show, as always.

I inevitably wind up thinking about the gap between the live event experience and reproductions on video or in VR. The gap is substantial, with the “shock and awe” of high powered, extreme dynamic range stadium effects, but there are still levers to pull.

A capability of the Quest VR hardware that would be fun to exploit is per-frame varying of the display persistence. The low-persistence display is only flashed “on” for about a tenth of each frame, so it doesn’t look blurred as you move your head around. This is a tradeoff between screen brightness and blur, which is just statically picked by Meta today, but it could be done dynamically on a per-frame basis (coordination would be tricky!).

Going to full persistence briefly would give an instant 10x brightness increase, which would be quite impactful on eyes adapted to the 10% brightness. You wouldn’t want to keep it there for long, because the blurring would be obnoxious and eyes would adapt to it, but it would be perfect for the epilepsy-inducing concert light flashes.

The clean way to do it would be with a 16 bit high dynamic range linear buffer that gets factored into an sRGB image for the display and a backlight flight time, but a hacky extra parameter latched with the frame submit would be the easiest thing to experiment with. Make it work, then make it clean!

A tech demo would be fun, but Beat Saber has a large enough user base that stepping up “concert lighting effects” might even be worth it from a net user value standpoint.

Trista: NIN - Friggin 🔥

Agents will be the biggest users of software. They’ll often need their own computers, identities, file systems, and tools to do their work. As a result, software will increasingly become API-first to be as useful to agents as they are to people. This is a huge opportunity.

Box: Agents need files to keep track of their work, they leverage files as context about the tasks they’re doing, and use them to share back and forth with their human counterparts. @levie spoke with @CNBC about and the importance of agents having their own filesystems.

Make agents work while you think, while you play, while you sleep.

This is Cursor Automations.

Cursor: We're introducing Cursor Automations to build always-on agents.

💥 GPT 5.4 is launching today! It's our best model ever, and it's also the most capable scientific model we've ever released.

GPT 5.4 Pro in particular is 🤯 based on early testing with scientists and mathematicians.

GPT-5.4 is launching, available now in the API and Codex and rolling out over the course of the day in ChatGPT.

It's much better at knowledge work and web search, and it has native computer use capabilities.

You can steer it mid-response, and it supports 1m tokens of context.

Codex app on Windows!

Andrew Ambrosino: The Codex app is now live on Windows.

The app runs both natively and in WSL, with integrated terminals for PowerShell, Command Prompt, Git Bash, or WSL.

We also built the first Windows-native agent sandbox — using OS-level controls to block filesystem writes outside your

Forgot to mention /fast!

I think people will like this.

Ahmed: Today we are introducing GPT-5.4 in codex. It's more token efficient and better at tool calling, computer use, and frontend development.

We are also introducing /fast to get a faster version of Codex.

Enjoy ❤️

We will be able to fix these three things!

Matt Shumer: I've been testing GPT-5.4 for the last week.

In short, it is the best model in the world, by far.

It's so good that it's the first model that makes the “which model should I use?” conversation feel almost over.

The biggest surprise: I barely use Pro anymore!

If you know me,

Model progress continues unabated, with GPT-5.4 showing significant improvements in critical knowledge worker tasks. In our Box AI tests, we saw a 6 point jump in agentic document processing, which is upstream from most automation workflows. GPT-5.4 is now available on Box.

Box: We tested @OpenAI's new GPT-5.4 model, and it showed a 78% overall extraction accuracy - up 6 points from GPT-5.2.

Our evaluation tested the model across real industries and document workflows:
→ Clinical data: +5 pts (81% → 86%)
→ Legal agreements: +3 pts (82% → 85%)
→

There was a nice time where researchers talked about various ideas quite openly on twitter. (before they disappeared into the gold mines :)).

My guess is that you can get quite far even in the current paradigm by introducing a number of memory ops as "tools" and throwing them into the mix in RL. E.g. current compaction and memory implementations are crappy, first, early examples that were somewhat bolted on, but both can be fairly easily generalized and made part of the optimization as just another tool during RL.

That said neither of these is fully satisfying because clearly people are capable of some weight-based updates (my personal suspicion - mostly during sleep). So there should be even more room for more exotic approaches for long-term memory that do change the weights, but exactly - the details are not obvious. This is a lot more exciting, but also more into the realm of research outside of the established prod stack.

Awni Hannun: I've been thinking a bit about continual learning recently, especially as it relates to long-running agents (and running a few toy experiments with MLX).

The status quo of prompt compaction coupled with recursive sub-agents is actually remarkably effective. Seems like we can go