Max Weinbach
@mweinbach
Analyst @creativestrat | Analyst and Market Research Firm | Typo ignorer
Email: max@creativestrategies.com
8K Following 307.1K Followers
Qwen 3.8 Flash Next updated numbers for M5 Ultra
We are now at 3740 tok/s prefill, 149 tok/s decode batched on M5 Ultra for Qwen 3.8 Flash Next
This is nearly double what it was yesterday
I have been using the Mac Studio with M5 Ultra for the past week, and it's the most powerful computer I've used.
Been a bit busy over the past week so my report will be slightly delayed, but just know... Qwen 3.8 Flash Next is the best model for the 256GB version, it rips
Show more
Oh yea so I have to say after a year of Ceramic Shield 2 on the 17 Pro Max and iPhone Air
This glass is magic it has 0 scratches and idk how that's even possible, it's just insanely good and I no longer worry about screen protectors or scratches at all
Show more
Mac Studio M5 Ultra
Backordered to... end of January
Apple’s C2 modem has support for mmWave and is in iPhone Duo
Price increase on the iPhone 18 Pro is… $100
Much lower than anyone was expecting
Let me give an example
It took a model compiled for the Snapdragon 8 Elite Gen 5 NPU, looked at the NPU drivers, traced the complication process to the hardware bindings for the end model, and got it back into essentially it's original weights with a 99.9% match on everything
Show more
While I'm not going to give any specific demos, part of the press release of GPT-6 Astra is that it can reverse engineer software binaries to understand logic without source code with high success rate
I can confirm and noticed it doing this to optimize/fix apps
Show more
Actually the craziest part of Astra is it will continue to work for days or weeks until something is done
I have it building a custom ROM for my Xiaomi 17 Ultra (the goal is to embed codex app-server) and it's been running for ~8 days straight and used over 1600 subagents
Show more
Another fun demo of Astra, I realized like 3.5 hours ago I didn't have a cool enough demo
In 75 minutes, I gave it the same simulate macOS 27 prompt. It's BY FAR the best output I've seen. If you log in, it also has cloud sync and working apps
Show more
I had early access to GPT-6 Astra and it's maybe the most insane model I've experienced
In Blender, I had it recreate Apple Park from just images. It did an absurd job.
ChatGPT Work did, in fact, book me a haircut for 90 minutes ago and it did work
I did get a haircut
@drbarnard I'd do 32GB on M6 I believe
I moved the @DiligenceStack agent evals over to Optima with Artificial Analysis. I'm about to run these models on it!
I can already tell you, this is a far better way to running these evals vs. what I was doing before
Show more
I'm really enjoying Gemini 3.7 Flash in Antigravity
It's really fast, seems to be really good, and the usage limits are insanely high
Grok Bot delegating to Cursor Cloud agents is pretty great
I’ve been trying this earlier, become a big fan. It’s also been good for handling emails and other work adjacent stuff. I’m happy with it.
Show more
I was using Daybreak Blue last night in Codex to do some work on finding vulnerabilities in a platform I'm building, it found like 5 pretty big ones and fixed them quickly, and ones that the other models didn't find. Pretty helpful.
Show more
Cook essentially just said on the earnings call that Chinese memory, if they could use it, would help on the supply side and they are not sure about the pricing side
Essentially, they just need access and it doesn't necessarily mean it's "cheap Chinese memory"
Show more
I love ChatGPT Work so much
