@ryanshrout · ranked 2 of 2 scored days
#26
– today
Thanks to Jason Andersen at Moor Insights & Strategy for taking a close look at @Signal_65 PINNACLE and writing up what he found, including the parts he thinks we still have to prove. The problem he starts from is the one I hear from every enterprise trying to move agents into
Adreno Matrix Cores put AI dedicated cores inside the graphics pipeline for the first time, so matrix acceleration now spans the NPU, the CPU, and the GPU on @Snapdragon. That is a notable move away from a decade of NPU TOPS being the number that mattered in mobile AI. Adreno
My inner hardware-geek just loves seeing stuff like this. The @Signal_65 PINNACLE benchmark getting some more hardware to exercise in the form of a full GB300 NVL72 rack!
DLSS 5 is shipping now, and I've spent a few days with it in NBA 2K27, the first game built to use it, running on a single RTX 5090. The community mods that have been floating around showed the raw capability of @NVIDIAGeForce neural rendering, but NBA 2K27 shows what it looks
A lot of hype around GPT-6 Astra this week. "It's true. All of it." On the enterprise agent work @Signal_65 PINNACLE measures, @OpenAI GPT-6 Astra makes about half the weighted errors of anything we had tested, finished 279 of 280 jobs end to end, and invented nothing. See
So, has OpenAI said anything about this on its platforms officially yet? This is actually a good reason to try out new GPT-6 or something even after I've been embedded in Claude for a while now. Has any model provider pushed back? Or created a US-specific version of the model?
An open-weight model now sits on the Signal65 PINNACLE cost frontier (https://t.co/mdZEANYF5y) above Opus 5 and GPT-5.6 Sol on agentic work, at 17 cents a correct task. Rent versus buy just got uncomfortable for the hosted tier. Also, Sol at max effort spent 3x the tokens for the
Meta may have just had its next Llama moment. Muse Spark 1.3 lands SECOND on @Signal_65 PINNACLE, 38% fewer weighted errors than Claude Opus 5, and a correct task costs $0.85 against $1.36 for Opus and $1.22 for GPT-5.6 Sol. Both of those just fell off the cost frontier. It