I am a IT Tech that love AI and video games!
Canada
Joined November 2013
- Tweets2K
- Following365
- Followers68
- Likes2.2K
Jérémie Dumont retweeted
Strata 0.2.0 is now live in beta! It includes its own in-app harness and terminal-serving interface.
github.com/Niko1221/Strata
A lot of performance improvements and support for low-end devices 👇
Codex speed check: Oct 2 → Oct 17 🚀
Fast mode (tok/s):
5.6 Sol: 73.6 → 105.9
6 Sol: 69.5 → 102.0
6.1 Sol: 29.3 → 68.8
6 Astra: 29.7 → 62.0
40 runs per date. Full Normal/Fast comparison in the video. @thsottiaux you work hard or add compute? or both?
Yeah we cannot just compare API equivalent for the sub usage, but it still hurt than my 20x sub will go doen to 10x only for the same price... I don't think 6.1-Sol only will make me keep the subs.
40 Codex tests: same 60K-token prompt, 5 runs per mode, medium reasoning, no tools.
Observed Fast gains:
GPT-6.1 Sol: +10.3%
GPT-6 Astra: +28.5%
GPT-6 Sol: +174.4%
GPT-5.6 Sol: +93.3%
GPT-6.1 Sol had the smallest Fast uplift in this sample.
@thsottiaux I think we need more compute on GPT-6.1 Sol!! We pay 2.5x for fast with usage, but only getting 10% more. Dots detroying Astra compute too...
Jérémie Dumont retweeted
NOUVEAU VOX POP!
La campagne électorale bat son plein. Je vous le confirme, la démocratie se porte exactement comme vous vous l’imaginez. 😂
Faites plaisir à votre entourage et partagez cette publication SVP. youtube.com/watch?v=2lxtyTus…
I don't see why people are mad about GPT-6-Sol?? I use it for my coding and it cost way less and do the job I ask. I use Astra for really task, but I main Sol right now. I can even use /fast mode with it and usage is fine.
If you're planning a PC build or upgrade, start paying attention to RAM and SSD prices now.
Intel's CEO said memory prices are up 5 to 7x and getting worse. He's raising CPU prices 10% in October. AMD is doing the same, 10% across GPUs and chipsets in Q4.
Samsung locked 70% of its memory capacity into HBM deals through 2031. SK hynix CEO says the shortage lasts through 2030. Hyperscaler capex is up 98% this year.
The data centers are eating the supply. Less commodity DRAM and NAND is left for your 32GB kit or 2TB NVMe.
Here's what that means for your wallet.
This isn't a bubble. The "AI slowdown" calls haven't slowed anything. Capex is up 98%. Nvidia's doubling shipments. SK hynix moved its cleanroom start 6 months earlier. Korea Eximbank raised its 2030 chip market forecast to $2 trillion.
The demand floor for memory is structurally higher than 2024. Not a spike. A new baseline.
32GB RAM + 2TB NVMe in the next 12 to 18 months is the most expensive window. Wait and you pay more.
I need to think about a use case for local AI model that run on my 5070ti. I'm using @MiaAI_lab github.com/MiaAI-Lab/Qwen3.8…
I benchmarked three 2-vCPU GitHub Actions runners against the exact same pinned CI workload: GitHub ubuntu-24.04, Blacksmith 2-vCPU, and Avrea 2-vCPU.
3 cold + 3 warm runs per provider, comparing Go dependency restore, npm install, Go tests, Go build, Docker build, cache save/restore, and total pipeline time.
Results, warm median:
⚡ Blacksmith: 82.48s
🟣 Avrea: 126.36s
⚪ GitHub: 166.33s
Blacksmith was:
• 2.02× faster than GitHub overall
• ~2.4× faster on the cold Docker build
• ~3.7× faster on cache restore
Avrea had the fastest cache restore at 1.39s, but Blacksmith won on total CI execution time.
Same workload. Same 2-vCPU class. Fresh runners. Median results.
#GitHubActions #DevOps #CI #CICD #Benchmark
I work on so many little projects for myself with AI. It's so awesome! But I really need a project I can share or even sell for needs that people have.
The Frontier said we need to slow down. I only need a good model that is fast and can code well. I think we need coding-only models back, maybe more specialized than a huge model that can do everything.
I don't know whit, so I can be wrong. What do you think?
OpenAI put the ChatGPT $200 subscription on pause to stabilize compute for existing users. Compute isn’t racking up quickly enough from demand after the Astra launch!
I put "never use Java" in AGENTS.md.
The model used Java.
Then it wrote a paragraph about why my trauma is not a valid coding standard.
AGI is here!
Astra is actually that good.
I’m on Low reasoning only and it still feels like a different class of model.
Problem: "efficient" for the task ≠ efficient for the subscription. Low still eats my usage alive.
Grok 4.6 is already in the same intelligence band for a fraction of the price. @elonmusk when 4.7 drops, can it do Astra-level work without the Astra-level meter?