mattwallace @mattwallace

techie, dad, author, inventor, perpetually curious; CTO & Builder all-in on #AI https://t.co/0MDUTxXXzb The Arena Joined January 2008

Tweets

7K
Followers

651
Following

248
Likes

1K

mattwallace @mattwallace

4 days ago

"having 6 working tests [ed note: of 12] is good progress". No, no its not.

0 0 0 31 0

mattwallace @mattwallace

4 days ago

so @COentrepreneur is like "I'm having 1 cup of coffee a day" and I'm like "yeah, me too"

1 0 0 53 0

Download Image

mattwallace @mattwallace

a week ago

ok, so I've got a lot of hw so I haven't done a lot of cpu inference, but wow. my aging intel i9-12900k dropping ~19tps with a qwen30b-a3 (2507) on a unsloth ud q5_k_xl is a pretty insane 🧠/hw ratio.

0 0 0 72 0

mattwallace @mattwallace

2 weeks ago

When I think about the 3d training data from nVidia empowering the gaming space and stuff like unreal, it feels like they went back in time and became their own grandfather.

0 0 0 77 0

Jay Hickey @jayhickey

2 weeks ago

@mattwallace @dsiroker It was the Chrome update, and it’s now fixed in the latest Rewind version!

1 1 1 68 1

mattwallace @mattwallace

3 weeks ago

saw databricks raise, was like "11 letters down, 15 to go!"

0 0 1 99 0

François Chollet @fchollet

3 weeks ago

Important point from Deep Learning with Python...

13 90 756 71K 300

Download Image

The biggest challenge for scaling #GenAI? Moving data. @KamiwazaAI tackles this challenge by processing data where it sits. With #IntelGaudi 3 + @IBMcloud, scale faster and automate more. Kamiwaza CEO @COentrepreneur explains how: intel.ly/3Us5OE8

1 3 4 1K 0

mattwallace @mattwallace

3 weeks ago

I forget which model I was watching reason recently, but I noticed it was thinking in shorthand. "Consider concurrency. Need to review docs. User may want more than 1 example." Honestly, if the SOTA providers are not using symbolic vocabs now I don't even know what.

Nous Research @NousResearch

3 weeks ago

35 61 437 67K 149

Download Image

0 0 2 177 0

mattwallace @mattwallace

3 weeks ago

This is such a crippling limitation of 5-pro. @sama This works perfectly fine in both of your major competitors. Your model is better, please don't hamstring the interface.

0 0 0 43 0

Download Image

mattwallace @mattwallace

3 weeks ago

What a time to be alive.

AK @_akhaliq

4 weeks ago

What a time to be alive.

13 27 249 46K 134

Download Video

0 0 0 44 0

mattwallace @mattwallace

4 weeks ago

Please stop breaking @github

0 0 0 38 0

mattwallace @mattwallace

4 weeks ago

So especially with GPT-5 out, I have to wonder - @kevinweil - why can I not have a `codex` or `gpt` cli I can attach a pro sub to so I can crank away with it? Why does the *desktop client* not have an mcp config, which would be ~more useful? Not enough GPUs to cover the uptick?