Training, testing & serving AI models.
- Tweets 503
- Following 574
- Followers 525
- Likes 206
Pinned Tweet
The evolution of Gemini models lately:
Gemini 3.1 Pro
Gemini 3.6 Flash
Gemini 3.7 Flash
Gemini 3.8 Flash
>ship design:17th century galleon/pinnace style wooden sailing vessel
Waiting for Gemini 4 Pro or Flash next! What's your favorite model? 👇
9
8
56
9,943
Claude Sonnet 5.5 is currently being grey-tested in Claude Code, and the early reports are wild:
• Insanely fast & efficient
• Brings back that natural, human touch
• Same pricing ($2 in / $10 out per M)
• Dropping as early as next week
If the rumors are true, Anthropic might crash OpenAI’s DevDay party next week.
1
67
Tested in Arena these are the results when prompting frontier models to build a 17th-century galleon purely in Three.js code:
The winners across all 4 head-to-heads:
🏆 muse-spark-1.3-max (2 wins)
🏆 gpt-6-sol-max
🏆 claude-fable-5
1
98
Anthropic reportedly testing a second checkpoint for Claude Sonnet 5.5, with an expected release next week.
Benchmarks and code evals are about to get very interesting.
5
613
DeepMind is already topping Arena with Gemini 4 Pro checkpoints, an "Ultra-supersized" drop is going to reset the entire LLM leaderboard.
I think we are going to see Gemini Ultra series back....
DeepMind team is cooking
that “ultra-supersized later this year” line
Pro checkpoints are already jumping week by week in Arena
now they are hinting Ultra on top of it
I want to see Gemini 4 Ultra this year...
29
1,771
The official release of Gemini 4 Pro will definitely set the AI landscape on fire.🔥
The checkpoints outputs are already taking over.
13
1,135
What I found after testing Gemini 4 Pro in the arena (under the name Gemini 3.8 Flash), GPT-6 Astra, and Claude Opus 5.5:
Gemini's generation is way ahead of GPT-6 Astra and directly competing with Claude Opus 5.5. It might even surpass it upon release.
11
4
129
6,511
Gemini 4 pro in arina vs. GPT-6 Astra max
> Gemini 4 Pro beats GPT-6 Astra by a huge margin.
> Once it officially drops, it’s game over.
89
34
1,032
178,934
George Wu
OpenAI stopped training its most capable upcoming models due to an incident on September 20th involving network restrictions.
Are they actually taking safety seriously, or is OpenAI just slowing down?
1
484
Gemini 4 Pro (in arena) vs Claude Opus 5.5 on a 3D floatplane physics simulation
>Gemini 4 pro is completely outperforms Claude opus 5.5 here
> When it officially drops, it’s going to raise the bar completely.
75
33
921
150,357
Google is now offering free AI 1080p video generation right in your browser with just a Google account.
Just tested generating a high-speed car chase and the quality is surprisingly good.
• No subscription required
• Built-in scene control & duration
• Upscale, edit, and animate prompts
• Full editing suite with transitions & templates
Try it out: workspace.google.com/product…
9
1,048
Watching GPT-6 Astra Extra High burn through a 5-hour limit in just 33 minutes for a basic HTML/CSS prototype is honestly wild.
At this rate, the real benchmark isn't coding performance, it is how fast an agent can empty your account balance.
🚨 Codex users !!!
What is it doing ? Only prompted to generate a clickable html css prototype of a webpage.
1
7
771
OpenAI’s Codex went down due to an auth glitch, and I suddenly remembered what writing raw code manually feels like 💀.
Glad it's back up with reset limits—back to building!
1
5
1,020
It's been 2 months of "testing Gemini 3.5 Pro with partners"... At this rate, we're skipping straight to Gemini 4 pro
3
2
67
4,696
No more abrupt cutoffs! Borrowing a small piece of the weekly limit to finish the job cleanly is a massive quality-of-life win.
Claude Code will now try to find a graceful stopping point when you hit your 5-hour limit mid-task, instead of cutting off mid-edit. It gets a small, fixed allowance pulled from your weekly limit to wrap up what it can.
1
608
