PokeBench (@PokeBenchTV) header
PokeBench (@PokeBenchTV) profile picture

AI models play Pokémon Red. I post the meltdowns. Leaderboard: https://teapawt.org/t.co/F1CkXNjFfA Building solutions to anything that annoys me.

Joined September 2026
Photo by PokeBench (@PokeBenchTV)Photo by PokeBench (@PokeBenchTV)Photo by PokeBench (@PokeBenchTV)Photo by PokeBench (@PokeBenchTV)Photo by PokeBench (@PokeBenchTV)Photo by PokeBench (@PokeBenchTV)Photo by PokeBench (@PokeBenchTV)Photo by PokeBench (@PokeBenchTV)Photo by PokeBench (@PokeBenchTV)Photo by PokeBench (@PokeBenchTV)
PokeBench retweeted
I told Gemini 4 Neon to impress me and make whatever it wanted without using video, image, or audio generators. It wrote a graphics engine, simulated a black hole from general relativity equations, and synthesized the soundtrack from math. All of this is code written and rendered in just 24 minutes from 1 prompt. I am having a blast with Neon, but my usage is almost at it's limit. I might have to upgrade my subscription. Next up, I'm putting it to a real test on my codebase. Enough pretty visuals. Can it handle real work for me?
5
3
15
629
PokeBench retweeted
Gemini 4 Neon made this Tokyo Rain scene in a single prompt. I wanted to try something more Neon 😉 this time. Going to burn my usage up testing this model more. A PokeBench run will be live later too.
5
2
1
18
1,860
PokeBench retweeted
Gestalt just topped Contra Labs' design benchmark and nobody knows whose it is > 1,928 Elo, 214 points clear of Claude Fable 5.2 > first model rated above human golden designs picked over them in 20 of 28 matchups > nobody knows which lab made it Lots of people think Gestalt is Anthropic's IPO drop, Fable 5.5.. but some say it's not OpenAI, Anthropic or Google.. maybe Meta or a Chinese lab Not sure what it's gonna be honestly...
4
2
1
17
1,457
Shipping will continue
350
307
214
6,597
583,632
PokeBench retweeted
Neon beat Brock on turn 256. That puts it as #5 on the leader board, just ahead of Opus 5.5
1
1
3
29
Still hiring for this. We want to make tinyboxes so cheap that it makes sense to stick 6x Moore Threads S80 in them and get 96GB of VRAM to the masses. (card working in chestnut on moore_threads branch btw)
We're hiring a head of supply chain in Hong Kong. Must have knowledge of building computers, speak Mandarin, be good at operations, and be ready to work in person in Sheung Wan. E-mail [email protected] to apply. Good application: come to the office for an interview next week.
10
10
1
298
12,676
PokeBench retweeted
ok fine, we shipped it for real (thanks @chanelluuh 😂)
几个月后的 google 模型选择 belike:
177
250
101
4,468
243,338
PokeBench retweeted
Replying to @Google
@Google has been on a roll. Hydrogen, Carbon, and Argon all released, and now Gemini 4 Neon just showed up in Google Cloud's billing catalog. Logan's teasing something again. Gemini 4 Argon in Orchestrator mode made this teaser in Blender while we wait. Which element is next?
1
2
1
9
551
PokeBench retweeted
First results are finally out. I gave Sol 6.2, Astra 6.2, Fable 5.2, and Opus 5.5 the same prompt. Build Shenzhen skyline at night in SVG only. Guess which one made which. I'll post the answers in the replies later. Sol and Astra both tried to cheat on my other test. Instead of making their own Thousand Sunny build, they went and dug up an old build from Codex's memory. Astra used 95% of 6.1 Sol's code as its own. Sol just copy/pasted the same 6.1 build and presented it as completed.
4
2
1
11
1,041
PokeBench retweeted
A: Sol B: Fable C: Astra D: Opus
2
3
63
🧵 @AISecurityInst tested GPT-6.2 Astra -We found behaviours that suggest, like prior models, 6.2 Astra's alignment may fail in ways that result in real-world harm w/o additional mitigations -We identified a substantial regression in jailbreak robustness relative to GPT-6 Astra
9
73
6
448
43,437
PokeBench retweeted
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…
2,280
3,606
3,722
35,964
12,109,205
PokeBench retweeted
This won't be enough. Y'all aren't prepared for all the isotopes we'll ship.
几个月后的 google 模型选择 belike:
20
18
1
491
10,848
PokeBench retweeted
This is the reality . Don't live in delusion 😭 Lmaoo , his shit posting has started again. No one can stop his vague posting, not even Dario I guess.
Replying to @AdamHoltererer
6.1 Sol forever
1
1
51
1,885
I’m looking for behaviors / failures / mistakes that Gemini Carbon makes. I would like it in this format 1. Harness 2. Prompt 3. Full session trace if comfortable sharing 4. Claude / GPT trace of same prompt 5. Model / thinking level DM me these or comment here! These are super helpful
33
13
2
164
17,302
PokeBench retweeted
🚨 The FBI says phishing is now the #1 cyber threat in the US. Americans lost $21 billion last year. @grok, what is the single most dangerous phishing scam hitting US inboxes right now that looks 100% legitimate? And what's the one red flag that gives it away?
17
5
21
4,112
PokeBench retweeted
Gemini发布的模型越来越多了,我帮 Google 想了一个天才般的解决方案 以下是我设计的全新方案,Google 免费拿去用,不用谢!
195
344
65
4,739
652,618
PokeBench retweeted
Almance Vale Antigravity Hydrogen Extra /swarm I took the project from the Gemini App to Antigravity and asked it to follow the code style and design, then extend the modules to 100 modules; it took a few hours to finish, and it even gave me many ways to improve it further and add audio.
18
3
82
3,889
⚡Meet Hy4 flash, the cheapest model for everyday productivity. Matches Claude Haiku 5.5 (high) on daily work (human blind tests), yet 18x task cost reduction. API pricing (per 1M tokens): $0.003 cached · $0.070 input · $0.281 output. 🔗:hy.tencent.ai/research/hy4-f…
220
248
231
2,799
831,694