David Sum

Tech Lead. AI. E-sports fan. Build now, think later.
Literally just came back to X lol

[Toronto Raptors @Raptors]

#7 FOREVER

[David Sum @davidsum]

Will GPT‑5.6 Sol on Cerebras at 750 tokens/s be available in Codex, or will it be API‑only? I assume it’s way more capacity‑constrained. If the current fast tier cost 2.5x more for a 1.5x speedup, I can only imagine the cost of this.

[David Sum @davidsum]

Given how popular Cursor is in companies, no one mentions that the team and enterprise plans charge an extra usage‑based fee per million tokens, regardless of token type. The cheaper the model, the higher the fee relative to the base cost.

Based on my own enterprise billing data, using GLM 5.2 High is ~60 % more expensive on a team/enterprise plan after this fee.

[jason @jxnlco]

advice to young engineers:

brother, I know you got that dog in you,

but you have to rest too.

In a world where you can do 10,000 things at the same time, you have to remember that rest is part of the work.

If you find yourself at 3:00 p.m on Thursday afternoon and you realize you've done all your work, take a breather. Go outside, stretch your arms out, and feel the sun on your face.

There will be some day in the future where you're going to have to work until 4:00am in the morning.

There will be some some day in the future where you know you need to push through the weekend.

But if you find yourself having a moment to breathe, just breathe.

We do not exist in a world that tells you that you cannot do anything anymore.

We only exist in a world where you are telling yourself to keep going.

You have to think of yourself as some kind of high-performance athlete.

You have to give yourself someone who, when they rolled their ankle in the third quarter, can push through.

We have to realize that those people have offseasons.

We have to realize those people invest in recovery.

But we live in a world where psychological capital feels more infinite than just the limitations of a biological body.

And you have to listen to your body.

[David Sum @davidsum]

Selecting specific models + reasoning levels is already confusing for most people. Plus some models being better at specific tasks vs others doesn't help. I think every harness will eventually need to have their own smarter model routing layer.

Optimizing for token usage, time, cost, quality across providers and usually within a single conversation sounds challenging. I think it would be a fun problem to work on.

[David Sum @davidsum]

What are the odds of Fable 5 being usage-based credits only? Hmm...

[Anthropic @AnthropicAI]

We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5.

We'll begin restoring access tomorrow, and will share an update soon.

We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.

[David Sum @davidsum]

AI use for devs is a level and constantly evolving playing field regardless of your previous engineering xp. Techniques and skills go from important to being solved within months.

Constantly trying to keep up is exhausting. Not too mention all of the overhyped ideas/repos that you need to filter through.

[David Sum @davidsum]

One of the best events in the CDL. EU sports/e-sports crowds are built different.

[Call of Duty League @CODLeague]

Pure Passion in Paris. 🔊

[David Sum @davidsum]

If your engineering team is still using non‑team / non‑enterprise plans for LLMs, then:

1. You likely don't care enough about observability, governance and helping your engineers understand token usage
2. You've yet to discover how difficult it is to manage usage‑based token billing in a team

[David Sum @davidsum]

OpenAI serving even a small % of its strongest model on Cerebras feels like a bigger deal than people realize. This was mostly for smaller/faster models up to now. 44GB SRAM, 21PB/s bandwidth. Exciting stuff and happy to see Cerebras used more.

[OpenAI @OpenAI]

Introducing a limited preview of GPT-5.6 Sol, our next generation frontier model, as well as GPT-5.6 Terra, a balanced model for efficient, everyday work, and GPT-5.6 Luna, a fast and affordable model for high-volume work.

https://t.co/OoM83SyISN

[David Sum @davidsum]

Is there an actual difference between GLM-5.2 providers that use NVFP4 vs FP8?

[David Sum @davidsum]

People trusted Headroom to reduce token usage and save money, but the Anthropic requests were breaking prompt caching for almost 7 weeks until a day or two ago (and still gained popularity). This issue showed 0 cache reads, 17/18 cache busts and ~1.15M bust-write tokens.

Headroom's own accounting still reported positive savings.

Anthropic cache reads are ~0.1x input cost, while cache writes cost more than normal input. People put too much trust into these projects.

https://t.co/sW0kwnDYnA

[Matt Pocock @mattpocockuk]

The "X technique reduces tokens by Y%" fad is so old

Can't believe people get taken in by this

[David Sum @davidsum]

@mattpocockuk Thank you for actually saying this... I'm all for people experimenting and trying out ideas, but so many of these trending projects are way overhyped and in a lot of cases even detrimental to their own stated goals.