Pinned Tweet
@OpenAI dropped their first Jev-style decision model, GPT-6 Luna Decision a couple of days ago. The benchmarks coming in say that it is currently the best Decision model on the market.
I put it to a test against the previous two leaders on the benchmarks, @typesafe's Jev and 7B in a 2v2 match of Battle City.
The yellow team are two independent instances of GPT-6 Luna Decision and the green team are Jev 1.13 and Decider 27B.
Out of 10 games, Jev and Decider win all 10/10 against GPT-6 Luna.
Setup:
- Every player has 3 lives and respawns at the initial spawn location after it is hit by a bullet.
- The game is computed in lockstep, and latency does not influence the outcome.
- The 10 games were played on 10 different maps.
I put it to a test against the previous two leaders on the benchmarks, @typesafe's Jev and 7B in a 2v2 match of Battle City.
The yellow team are two independent instances of GPT-6 Luna Decision and the green team are Jev 1.13 and Decider 27B.
Out of 10 games, Jev and Decider win all 10/10 against GPT-6 Luna.
Setup:
- Every player has 3 lives and respawns at the initial spawn location after it is hit by a bullet.
- The game is computed in lockstep, and latency does not influence the outcome.
- The 10 games were played on 10 different maps.
1
0
2
164
Been coding locally with Qwen 3.8 27B, reminds me of coding with Opus 4.6 earlier this year.
Running Opus 5.5 on another project at the same time.
Since February we got significantly better raw capability, and a model that would be SOTA eight months ago now runs on one GPU.
Running Opus 5.5 on another project at the same time.
Since February we got significantly better raw capability, and a model that would be SOTA eight months ago now runs on one GPU.
Introducing Claude Opus 4.6. Our smartest model got an upgrade.
Opus 4.6 plans more carefully, sustains agentic tasks for longer, operates reliably in massive codebases, and catches its own mistakes.
It’s also our first Opus-class model with 1M token context in beta.
0
0
3
155
One reasoning model vs. three decision models in my Battle City inspired game.
The goal: see how a System Two model (slow, deliberate reasoning) differs in capabilities and behavior from lightweight System One decision models (fast, reactive choices).
Result: out of 5 games on pected, Sonnet won all 5 games.
The goal: see how a System Two model (slow, deliberate reasoning) differs in capabilities and behavior from lightweight System One decision models (fast, reactive choices).
Result: out of 5 games on pected, Sonnet won all 5 games.
3
1
4
110
To be clear, this is a demonstration of how differently these models approach the same problem, not a benchmark.
1
0
1
10
Setup:
- Every player has three lives and respawns at its initial spawn location when hit. Friendly fire is disabled, and tanks within the alliance count as friendlies.
- The game runs in lockstep, so latency has no effect on the outcome. It is only recorded.
- All models receive the same input: the map state encoded in ASCII, the game rules, and additional game-state descriptions.
- Sonnet outputs a single tool call containing the actions for both tanks at each step.
- Every player has three lives and respawns at its initial spawn location when hit. Friendly fire is disabled, and tanks within the alliance count as friendlies.
- The game runs in lockstep, so latency has no effect on the outcome. It is only recorded.
- All models receive the same input: the map state encoded in ASCII, the game rules, and additional game-state descriptions.
- Sonnet outputs a single tool call containing the actions for both tanks at each step.
0
0
1
10
Done with @OpenRouter.
Too many low-quality providers now. Filtering them all out is more work than just pointing at the official APIs + the few inference providers that are actually consistent.
Too many low-quality providers now. Filtering them all out is more work than just pointing at the official APIs + the few inference providers that are actually consistent.
0
0
1
66
Man, are these decision models bad at planning and spatial reasoning.
Here's TypeSafe's Jev fighting Perplexity's Decider in a Battle City knockoff.
Next is whether any expertise can be distilled into them, and what a dual-process agent looks like in the same match.
Here's TypeSafe's Jev fighting Perplexity's Decider in a Battle City knockoff.
Next is whether any expertise can be distilled into them, and what a dual-process agent looks like in the same match.
0
0
2
81
Opus 5.5 seems to be super token efficient and the benchmarks aren't lying about performance.
@AnthropicAI outdid themselves this time.
@AnthropicAI outdid themselves this time.
0
0
1
83
Maneuver yourself into a massive clusterfuck — the biggest economic bubble the world has ever seen — then immediately start crying about safety. A move straight out of the European Union playbook…
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must-p…
0
0
5
343
Seems that AA had to come up with a hotfix to avoid further backlash from the OpenAI propaganda release (i.e. GPT-6 Astra)
Spoiler alert! Fable is still at the top…
Spoiler alert! Fable is still at the top…
Anyone else thinks that GPT-6 is just a propaganda operation to secure that $500B loan for OpenAI?
On a separate note, where can one short AI? Asking for a friend...
0
0
2
152
Anyone else thinks that GPT-6 is just a propaganda operation to secure that $500B loan for OpenAI?
On a separate note, where can one short AI? Asking for a friend...
On a separate note, where can one short AI? Asking for a friend...
2
1
7
529
Anes Valentic retweeted
For my first post, I’m sharing a letter @nvidia signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
rontier closed models and frontier open models.
images.nvidia.com/pdf/Open-Weigh…
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
rontier closed models and frontier open models.
images.nvidia.com/pdf/Open-Weigh…
15,982
29,270
171,455
66,501,137
I'm just going to say, try Kimi K2.6 if you haven't already. You won't regret it.
📢 Kimi K2.6 API is live
• Input Price (Cache Hit): $0.16 / M tokens
• Input Price (Cache Miss): $0.95 / M tokens
• Output: $4.00 / M tokens
Kimi K2.6 is our latest + most intelligent model - stronger long-horizon coding, better instruction following & self-correction.
/image/video), thinking + non-thinking, 256K context. Supports tool calls, JSON / Partial mode, and web search.
Try it → platform.kimi.ai
0
0
1
137