Pekka Lund

I'm a Gemini Pro subscriber and I have lost my Gemini 3 Pro GA. Can you provide me a new one and how soon?

Aika monimutkaista ja työlästä tuollainen kokeilu kuitenkin olisi, ja ottaisi aikaa ennenkuin siitä saa kunnon käsityksen.

Mutta onneksi mulla on ratkaisu. Hankit sen kämpän Espanjasta, jostain läheltä rantaa, ja mä voin koekäyttää sen oikein huolella puolestasi.

Could be coherent for a change.

Can anyone now claim this isn't true?

"because "humans suck and deserve to be replaced.""

Also, risks private capture? How would that be different from what the US has now?

Claude Opus 4.6 (Adaptive) - Intelligence, Performance & Price Analysis Analysis of Anthropic's Claude Opus 4.6 (Adaptive Reasoning) and comparison to other AI models across key metrics including quality, price, performance (tokens per second & time to first token), conte...

Artificial Analysis details here the amounts of answer & thinking tokens models have used when running their benchmark. 4.6 (Adaptive) used twice as much as 4.5 (Thinking).

It presumably costs them in using their GPUs less efficiently. Hard to say what the real multiplier is though.

Speed up responses with fast mode - Claude Code Docs Get faster Opus 4.6 responses in Claude Code by toggling fast mode.

Anthropic themselves state it's about API configuration (so not hardware or quantization):

"Fast mode is not a different model. It uses the same Opus 4.6 with a different API configuration that prioritizes speed over cost efficiency."

We can of course grant respect for many reasons but our mortality in itself surely isn't an achievement to be proud of.

I tried problem 15 with Pony Alpha. It failed.

It took a few tries to get the request through and then it just stopped mid-reasoning a couple of times. But eventually I got an (incorrect) answer.

OpenRouter recommended adding my own API key for (Stealth) company for avoiding rate limit errors.

I find it distasteful that people believe people do in any other sense than the bots do.

This is not what is happening at all.

The amount of misinformation on BlueSky about AI is insane, and it keeps promising that AI is all hype that is going away soon.

A really dangerous position that cedes all AI policy and decisions about how it will be used to others.

Also Futurism is clickbait

I happened to get responses from two models that were both named Gemini 3 Pro with the Google logo. 🤷‍♂️

One failed, one succeeded.

People have claimed that Gemini 3 Pro GA is currently being tested in arena.ai Battles and should be possible to identify by missing the Google logo in front of the model name.

So I tried problem 15 there, which Gemini 3 Pro had failed.

matharena.ai has tested several models with the freshly released AIME 2026 I problems. All of them are now so good in math that AIME doesn't really work as a benchmark anymore.

Only 3 models failed to solve one problem with any of their 4 tries. GPT-5.2 only failed 1/4 tries on 2 problems.

Psychiatry in 2026:

What's wrong with everybody?

- No, what's wrong with you?

Are you saying I'm the crazy one?

- No, that's the problem. You should be!

It's pretty impressive if it does so well without any. But maybe it just reasons almost as much in final outputs instead of separate <thinking> tags.

I guess it should be easy to test by forcing it to provide final answers only. It should perform pretty poorly if it doesn't have hidden reasoning.

Artificial Analysis classifies the base model as a non-reasoning model. Apparently it doesn't generate <thinking> blocks then, which Anthropic calls extended thinking. But I don't know if it still does something like that behind the scenes.

Adaptive thinking Let Claude dynamically decide when and how much to think with adaptive thinking mode.

Adaptive means it decides reasoning effort dynamically. Presumably scores could be somewhat better if max reasoning was used always, as with GPT-5.2 (xhigh).

At least one Z.ai employee has indicated GLM-5 will be released in February. Timing would now fit with release before Chinese New Year.

Description given by OpenRouter is similar to how Z.ai described GLM-4.7. Context window matches that or Claude. But clearly not claude as it said about ASI: "Would it preserve humans? Only if we're instrumentally useful".

Tells now about Tiananmen but others have reported refusals before.

GLM 5 seems to be the common guess.

Doesn't feel like Grok, likely too early for Meta.

I think you mixed up models.

That sounds like a Grok defense.

Or Pro

Oh, wow, Claude Opus 4.6 took the top spot in Artificial Analysis Intelligence index.

Gemini 3 Pro, which used to be in the lead (before the benchmark was updated), is now fifth. And they haven't even tested GPT-5.3 Codex yet.

Epoch now made a thread about it.

Opus math skills have increased A LOT in the 2.5 months between 4.5 and 4.6.

They now scored 10/48 = 20.8% in Tier 4 and are in second place behind GPT-5.2 Pro.

Gemini 3 Pro GA/Deep Think, where are you? Time to show what you can do!

I think ARC-AGI-2 illustrates the situation nicely. Many are now surprised LLMs exceeded human average baselines so fast. But the raw intelligence was already there long time ago. It just needed to be unlocked by scaling up the capabilities for tracking many small details in the context and such.

My feeling is that, of all the AI lab leaders, Amodei is probably the most direct in saying what he actually believes.

But all of them need to be careful what they say. They need to consider marketing, funding, regulation, and not to freak people out.

GPT-5 lowers the cost of cell-free protein synthesis An autonomous lab combining OpenAI’s GPT-5 with Ginkgo Bioworks’ cloud automation cut cell-free protein synthesis costs by 40% through closed-loop experimentation.

For physical work, the next step seems to be automated labs, like in this case OpenAI just published. Altman recently stated they are going "back and forth a lot about whether we should be building automated wet labs for every field".

General purpose (humanoid) robots comes after that.

It's harder to predict when that would lead to fast takeoff or accelerating beyond our control, as it depends e.g. on availability of computing power, political decisions and other hard to predict developments.

So I think the smarts for that are there much sooner than other necessary resources.