Joined August 2023
- Tweets 1,336
- Following 281
- Followers 267
- Likes 6,318
224 Photos and videos
Pinned Tweet
May 5
Make it you against the matrix. Winning is the only way.
6
164
Jun 12
The latest SOTA image model from OpenAI for creating, generating & editing images.
Try it now 👇. Access 50 frontier models too.
commonstack.ai/account/signu…
Jun 12
GPT Image 2 is available now on @commonstack_ai
This is the latest SOTA image model from OpenAI for you to create, generate or edit images.
Try now on Commonstack 🎨
17
134
Jun 10
Claude Fable 5 took #1 on the Artificial Analysis Intelligence Index with 64.9—nearly 5 points clear of GPT-5.5. Anthropic now owns the top 2 spots.
It’s a Mythos-class model with extra guardrails: flags risky queries in cyber, bio, chem.
2
8
98
Jun 10
Runs at frontier level on agentic tasks, coding, and hard science benchmarks. 1M context, $10/$50 per million tokens.
Big leap if you need serious reasoning.
3
42
Jun 9
Scoring 55 on the Artificial Analysis Index, it sits right alongside the top closed-source models. What makes M3 different? It's the first to combine frontier coding, native multimodality, and a massive 1M token context window in an upcoming open-weights release
1
12
146
Jun 9
This is a massive win for agentic workflows. With the weights dropping soon, the "open" landscape is about to get a whole lot more capable. Experience open models with @tryParallax iykyk
1
5
36
Jun 8
NVIDIA Nemotron 3 Ultra was evaluated by Artificial Analysis on Terminal-Bench v2.1 under restricted turn-usage limits using a customized Terminus 2 harness. Each model received a defined turn limit, with results tracked across four increasing limits.
2
11
213
Jun 8
Under these conditions, Nemotron 3 Ultra recorded the lowest time per task across all turn limits while achieving competitive task completion rates, positioning it on the Pareto frontier of performance versus latency in this evaluation. The model is an open-weight release.
1
3
37
Jun 8
Build with @tryParallax to have the best experience of access local ai and participate in intelligence as we crave more inteligence.
3
19
Jun 4
Harvard just validated something the async RL community has argued for a while- The standard LLM pipeline—Pre-train → SFT → RL sequentially—is suboptimal.
Three findings from Harvard that matter:
Jun 4
arxiv.org/abs/2606.04272
How does applying RL after pretraining compare to using intermediate SFT? The most interesting effect would be that sharpening of distribution does not happen with direct RL. (Though scale and effect are small, and Dolmino mix itself resembles SFT data.)
1
10
548
Jun 4
Sequential pipelines create artificial bottlenecks. Harvard’s research validates the core assumption of async RL: you don’t need perfectly synchronized, on-policy data to get great results.
Async RL workers can continuously generate rollouts using slightly stale policies
1
3
25
Jun 4
without sacrificing quality.
Paper: arxiv.org/abs/2606.04272
Echo-2 from @Gradient_HQ builds on this directly: async RL with bounded staleness, running continuously—not sequentially: gradient.network/blog/echo-2…
3
28
Jun 2
Recent @Meta new ScaleAcross Explorer research paper tackles the immense complexity of Scaling AI training beyond single buildings to "scale-across" architectures via distributed training. As we push toward frontier models, the system design space explodes as we've seen recently
1
10
128
Jun 2
Just as we’ve seen with efforts like @Gradient_HQ ECHO-2 in optimizing distributed training, ECHO-2 proves the payoff—delivering 10.6× cost reductions in RL post-training ($4,490 → $425 for a 30B model) while preserving quality and accelerating training up to 13×.
1
3
59
Jun 2
the future of AI isn't just bigger clusters— The future of frontier training isn’t bigger homogeneous clusters. It’s intelligent orchestration of whatever silicon is available.
Read more on echo-2: gradient.network/blog/echo-2…
1
20
Jun 1
NVIDIA’s RTX Spark is pushing hard into local AI.
Parallax makes it vendor-independent: cluster RTX Sparks, 5090s, Macs — whatever you have — and run your open models pipeline-parallel across them.
Fully private. Local. Maximum hardware utilization.
Your AI, your cluster. âš¡
Jun 1
nvidia going all in on local ai.
here's our take: it shouldn't depend on which chip you bought.
sparks, macs, the 5090 already on your desk, we cluster across all of it and split your favorite model pipeline-parallel so it runs fully private and local.
10
138
Pascal2_22./ retweeted
May 30
🚨 Arne Slot and Liverpool, it’s over. Andoni Iraola clear favorite to replace him. 💣🔃
2,663
5,596
59,004
4,952,434