Google announces Gemini 3.1 Pro, says it’s better at complex problem-solving
- Thread starter JournalBot
- Start date
Is there any practical difference between the two designations (preview/GA)? In our enterprise environment (GitHub Copilot) Gemini 3.0 Pro (preview) was enabled long time ago. I see no difference between using it and Claude models (apart from the merits of the models). Could it be that different LLM providers simply label the models differently?
Upvote
1
(5
/
-4)
While your logic has some validity to it, LLMs have a point too. They can't read your mind (not yet). It is likely that there are people that like this LLM/agent behavior.
Upvote
-16
(0
/
-16)
Google Ads rankings have been based on machine learning for decades. The only operative part of your post was "to my knowledge".
Upvote
-7
(3
/
-10)
Is that a serious question? Personally, as a software designer, I almost stopped doing the traditional coding. I create the specs and then ask LLMs to do the coding. I am a full time employee but there are many freelance software developers who pay hundred(s) of dollars out of pocket monthly for tools like Claude Code because the boost in productivity easily pays for it.
Upvote
-1
(13
/
-14)
Concur. I am a professional software developer with 30 years of industry experience. I now use LLMs for research, rubber ducking, and as IDE plugins to generate code. I wouldn't consider going back to not having them.
Outside of work they are excellent for analyzing scans of handwritten documents and for searching archives that I uploaded to Google Drive. I have one Drive folder that contains the entire City of Berkeley Clerk's archives and Gemini can search it better than any classical retrieval system could hope to do.
Upvote
17
(18
/
-1)
One day after they finish with this AI fad I hope they can find to time to fix gmail so I don't get 2+ spam emails a day that land in my proper inbox and alert me on my phone via push notification.
No amount of report spam and block makes any difference.
No amount of report spam and block makes any difference.
Upvote
7
(7
/
0)
I have installed GH Copilot CLI on my home computers (Windows and Linux) and use it for many things: maintenance, troubleshooting, configuring Docker containers etc. It helps tremendously.
Upvote
3
(6
/
-3)
Upvote
2
(3
/
-1)
So, if you get 50% on a benchmark, does that mean that the AI model is gaslighting you the other 50%
Upvote
-1
(2
/
-3)
apple to oranges. They had to compare to ChatGPT 5.2 pro (extended), not thinking models. I found gpt 5.2 Pro to be remarkable and accurate.
Upvote
-1
(0
/
-1)
Oh boy, I guess they're going to make 3.1 Pro super effective for a couple days to ride this wave of news coverage, and then dial it back to "headless chicken" afterward.
Upvote
1
(1
/
0)
sigmasirrus
Ars Scholae Palatinae
Machine learning =/= LLMs. The former is wonderful, useful, and not killing the planet.
Upvote
10
(12
/
-2)
This sounds awfully familiar. As if there is a certain person in a high place of power who follows this programming.
Upvote
5
(5
/
0)
A friend of mine authored an excellent paper on a generalized version of that, "proxy failure."
https://pubmed.ncbi.nlm.nih.gov/37357710/
https://www.researchgate.net/profil...an-inherent-risk-in-goal-oriented-systems.pdf
Upvote
7
(7
/
0)
They just don't get it, do they? Computers are supposed to be deterministic. It's in the very name: they compute, that's a deterministic process. I know benchmarks are used routinely in machine learning, although academics getting published because they bumped the accuracy a few percentage points is quite a different thing from me having to tell my users the algorithm is now slightly less likely to produce incorrect results. In practice, how often can we accept even 5% error rate? Who wants to have to go back and check because we just can't be sure the computer's outcome is accurate?
And that's machine learning, where to problem is usually quite circumscribed. Applying the same logic to these magic oracles that are sold as being capable of doing almost anything makes it seems even harder to trust them. With something that powerful, you're almost talking about something generic like a CPU. Who would accept a CPU that doesn't do exactly what it should even 0.1% of the time? And I have yet to see any LLM come close to that. Given that, who cares if they have benchmarks with number going up?
And that's machine learning, where to problem is usually quite circumscribed. Applying the same logic to these magic oracles that are sold as being capable of doing almost anything makes it seems even harder to trust them. With something that powerful, you're almost talking about something generic like a CPU. Who would accept a CPU that doesn't do exactly what it should even 0.1% of the time? And I have yet to see any LLM come close to that. Given that, who cares if they have benchmarks with number going up?
Upvote
3
(6
/
-3)
This is fucking hilarious.
Thank you for confirming that those of us with "analyst" in our job titles probably shouldn't worry too much about AI.
Upvote
9
(13
/
-4)
That's a lovely bit of sycophancy. I hope you are checking your business plans with humans, too.
All the titles and headings at the top also make clear that you have (deliberately or unintentionally) made very clear to the AI what you want, so of course it is happy to deliver. That's what they do.
Upvote
10
(12
/
-2)
Just a heads up, ChatGPT et. al. are not mechanical turks. The mechanical turk was the first "computerized chessboard" that was claimed to be operated independently as a "thinking machine". Under the board was a human being making the decisions and moving the pieces by hand. If ChatGPT and the likes were mechanical turks, there would be someone on the other side answering the questions - there isn't.
Upvote
4
(5
/
-1)
I didn't watch the entire video, but AI (a group of technologies that's been around since the 1960's) and GenAI/LLMs (ChatGPT, et. al., since 2020) are two different subjects. I wouldn't be surprised at all if AI, in the broad term, can help in mathematics. LLMs doing Maths? Never going to happen.
Upvote
6
(7
/
-1)
Okay, I will grant you it's potentially good for software development purposes (but I have my doubts, but for this post I'll accept the claim).
Anthorophic lost 3 billion dollars in 2025. They made 1 billion dollars in 2024. What will you do once Anthorophic no longer is in business?
Upvote
7
(7
/
0)
How can an LLM, out of all things, give driving directions, and why would you trust it? Google has other products that are better for directions (Google Maps, for example)
Upvote
1
(3
/
-2)
If I was a betting kind of girl, I'd speculate that the poster is experiencing AI psychosis.
Going to be seeing a lot more of that in the months ahead.
Also, making multiple accounts to get around new-user post limits - very nice!
Upvote
10
(12
/
-2)
Seriously? an AI for maths would be amazing, being able to solve a problem using a chat interface where it selects appropriate formulas, tells you the required inputs, crunches the numbers into an answer, and then explains the answer would be a massive improvement for a huge number of people.
Personally I find it better to ask it to write a python program i can at least check for correctness, but many people really want it to just spit out an answer. Intuitively you expect computers to be accurate at computing things even though this particular technology doesnt get it without giving it tool use.
Upvote
-3
(0
/
-3)
Switch to Gemini?
In all seriousness, Uber ran at a loss for 14 years and is still in business. Amazon, for 9. The profitability of the early years of a business is really not a good indication of how long it's going to be around.
Upvote
-3
(2
/
-5)
THIS is a perfect example of the "lots of words, no value" output that LLMs are excellent at producing.
How so? How isn't it a hallucination? How isn't this fiction. Why isn't this expanded on?
I never heard of Sovereign Stack, but I heard of Sovereign Cloud Stack, which is a European Vender Neutral Cloud Ecosystem. See https://sovereigncloudstack.org/en/
I never heard of this. A Google search shows It appears to be something about a Quantum Sharing Technology. See https://journals.aps.org/prapplied/abstract/10.1103/PhysRevApplied.22.044018 . You can't assume people know what you're talking about.
Who invented these "mathematically verified architectures"? If it's Sovereign Stack and Coherent State Network Protocol, they aren't architectures. Who verified the architectures mentioned above? How were they verified? Why isn't a brief overview of these architectures mentioned.
So the Sovereign Stack and the Coherent State Network Protocol are technologies you invented? Who verified these technologies? How would a normal reader supposed to know you invented them.
Looking at Microsoft's AutoGen Github with comments by a person similar to your name, it appears the commits haven't been approved by Microsoft.
What is "it"? An Operating System is a specific type of software that controls the low level hardware of a computer system.
This reads like AI psychosis. Please get help.
Upvote
10
(12
/
-2)
Amazon and Uber are bad examples. For Amazon and Uber, the loses were based on infrastructure and lowering prices to attract customers. Anthoropic's (and OpenAI) losses are caused by training. Training, by definition (for AI) will never end.
Early year losses, by itself, may not be a good indicator of success, however how they are losing money is a good indicator of success.
Upvote
3
(5
/
-2)
The more I (try) to read his post, the more I think it's a troll, not psychosis.If he really believed what the LLM is generating, he wouldn't include the background information, but only the "Kinetic Reply". But it really is creepy that his LLM is outputting military/ police jargon as part of a "psychoanalysis" (counterstrike? Kinetic replies?).
Upvote
1
(2
/
-1)
No - the default Gemini Enterprise model is the 2.5 flash (non thinking) one which is very outdated now.
Upvote
0
(0
/
0)
It always depends on your specific usecases.
Certain lower end models can perform perfectly well on your usecases but not on others.
The higher end models are just more likely to handle complex usecases better.
Upvote
0
(0
/
0)
According to the people here, people like Terence Tao/Andrew Ng/Hinton are sell outs so you'll just get downvoted or, worse, ignored.
Upvote
-4
(1
/
-5)
I was laughing through most of it, but I'm kinds tempted to point out at the end here that your AI bot is so bereft of context that it seems ignorant of the fact that it sounds like it's about to launch into a manifesto that's a few degrees short of sedition, if chatbots were capable of sedition.
You imbeciles are going to create whole new categories of regulatory law and a series of neverending debates that will finally answer the question "what if the Talmud, but on Social Media?"
Upvote
4
(5
/
-1)
Yeah, ummm, I'm also going to assume it's someone very new to the forum. Remember, that post also explicitly talks about my job.
"Kinetic replies" to me, regarding my job, is a capital offense, this is not a joke and seriously not something to joke about. This is clearly an LLM model, but I would like people to consider exactly how close that LLM bot came to making explicit threats against a government employee as part of some effort towards some "Sovereign" conspiracy.
Again, I want to emphasize that I am laughing at the stupidity of this, and I must reiterate my full and vocal support of the first amendment. But this shit is going to get out of hand because LLMs lack mens rea and equally importantly they lack the social cues that most of us learn about not joking about certain topics (like delivering "real world payloads" to people on internet fora, jeez!).
Upvote
1
(2
/
-1)
You know, the "real world payload" thing made me realize the full scale of this madness: we're spending billions to create artificially unintelligent trolls, and yet no one is spending even a fraction of that on the one real-world payload that should be an expected consequence of internet banter: beer.
For a mere hundred million dollars, I will gladly found a startup to allow people to push a button to buy someone a beer over the internet for making a good forum post.
For a mere hundred million dollars, I will gladly found a startup to allow people to push a button to buy someone a beer over the internet for making a good forum post.
Upvote
6
(6
/
0)
macosandlinux
Ars Praefectus
Ok, 'google for me' XYZ and find only XYZ and not clickbait, AI slop or ads.
Can't do it? Oh, that's bad. Maybe 3.2 Pro will.
Upvote
2
(2
/
0)
macosandlinux
Ars Praefectus
Ryan's been on the Google beat, so this might just be that, or he covers for now AI @ Google? Still waiting for a Ars response on the ethical breach. Is there one anywhere?
Upvote
5
(5
/
0)
No response about it. I doubt they would make a public statement about it either.
Upvote
3
(3
/
0)
If it's a bad comparison, it's because anthropic can shed expenses way easier than the others. Because you're wrong - training is not required to go on. They're only doing that for competition and growth.
LLMs will go through the same evolution as all tech these days and wind up enshittified. And when that happens, AI companies will lay off most of their AI researchers. They could stop training and coast on their existing models. Likely, they'll just focus on training low cost models that are optimized for cheap serving.
And in doing so, they'll drop their training related costs to near 0 and shareholders will get their dividends.
Last edited:
Upvote
-2
(2
/
-4)
How large are the software projects you work on? And what language(s)?
I'm genuinely curious. When I see people claiming wonders it seems to usually be a hobbyist, academic/researcher, or tech executive - folks that work on small projects or scripts and/or don't actually have a software background. Meanwhile I'm in hundreds of thousands of lines of 20-30 year old C/C++ code at work (orders of magnitude beyond what an LLM can handle), and at home Gemini Pro struggles to generate simple working snippets of elisp code for Emacs half the time. So there's a big disconnect from my perspective.
Disclaimer: I think LLMs are cool and useful. For me they serve the same use cases as a search engine.
Upvote
5
(7
/
-2)
It is hilarious, and I am disappointed that the original has been disappeared. [I wouldn't mind if 'MohskiBroski' had just gotten the permaban he no doubt anticipated, but I'd prefer if the original had been left in all its glory for us to ponder whether it is satire or not. In my opinion it is - good satire strikes at all.]
[EDIT: heh, he's real. Still funny, though.]
Last edited:
Upvote
-2
(2
/
-4)