503 comments

[ 4.4 ms ] story [ 331 ms ] thread
(comment deleted)
(comment deleted)
Isn't every new AI model the "most <adjective>"?

Nobody is going to say "Announcing Foobar 7.1 - not our best!"

gobble 2.0 - a bit of a turkey
GPT-4.5's announcement was the equivalent of that.

"It beats all the benchmarks...but you really really don't want to use it."

They even priced it so people would avoid using it. GPT-4.5's entire function was to be the anchor of keeping OpenAI in the news, to keep up the perception of releasing quickly.
My assumption was that the pricing was because it really was that expensive for whatever reason. I'm keeping fingers crossed that they're going to do some kind of 4.5 mini at some point that will be more affordable.
Except for GPT 4.5 and Claude 3.7 :/
You're not wrong, but that just means the <adjective> is where the bulk of information resides. The trade-off matters. Maybe it's a model with good enough quality but really cheap to serve. Maybe it's a model that only plays poker really well but sucks at everything else because it bluffs too much. Etc. etc.
Sure but that adjective matters. Could be cheapest, "intelligent", fastest, etc... it's rarely all three of them.
Stable Diffusion definitely had a few of those.
Same with new phones. The new phone is always the fastest cpu, gpu, and best camera ever!
why not enable Canvas for this model on Gemini.google.com? Arguably the weakest link of Canvas is the terrible code that Gemini 2.0 Flash writes for Canvas to run..
I'm guessing it should be enabled eventually. @logankilpatrick thoughts?
Slight tangent: Interesting that they use o3-mini as the comparison rather than o1.

I've been using o1 almost exclusively for the past couple months and have been impressed to the point where I don't feel the need to "upgrade" for a better model.

Are there benchmarks showing o3-mini performing better than o1?

I find o3 at least faster to get to the response I care about, anecdotally.
I noticed this too, I have used both o1 and o3 mini extensively, and I have ran many tests on my own problems and o1 solves one of my hardest prompts quite reliably but o3 is very inconsistent. So from my anecdotal experience o1 is a superior model in terms of capability.

The fact they would exclude it from their benchmarks seems biased/desperate and makes me trust them less. They probably thought it was clever to leave o1 out, something like "o3 is the newest model lets just compare against that", but I think for anyone paying attention that decision will backfire.

Probably because It is more similar to o3 in terms of size/parameters as well as price (although I would expect this to be at least half price)
It's a reasonable comparison given it'll likely be priced similarly to o3-mini. I find o1 to be strictly better than o3-mini, but still use o3-mini for the majority of my agentic workflow because o1 is so much more expensive.
Why would you compare against all the models from a competitor. You take their latest one that you can test. Openai or anthropoc don’t compare against the whole gemini family.
The benchmark numbers don't really mean anything -- Google says that Gemini 2.5 Pro has an AIME score of 86.7 which beats o3-mini's score of 86.5, but OpenAI's announcement post [1] said that o3-mini-high has a score of 87.3 which Gemini 2.5 would lose to. The chart says "All numbers are sourced from providers' self-reported numbers" but the only mention of o3-mini having a score of 86.5 I could find was from this other source [2]

[1] https://openai.com/index/openai-o3-mini/ [2] https://www.vals.ai/benchmarks/aime-2025-03-24

You just have to use the models yourself and see. In my experience o3-mini is much worse than o1.

I wonder what about this one gets the +0.5 to the name. IIRC the 2.0 model isn’t particularly old yet. Is it purely marketing, does it represent new model structure, iteratively more training data over the base 2.0, new serving infrastructure, etc?

I’ve always found the use of the *.5 naming kinda silly when it became a thing. When OpenAI released 3.5, they said they already had 4 underway at the time, they were just tweaking 3 be better for ChatGPT. It felt like a scrappy startup name, and now it’s spread across the industry. Anthropic naming their models Sonnet 3, 3.5, 3.5 (new), 3.7 felt like the worst offender of this naming scheme.

I’m a much bigger fan of semver (not skipping to .5 though), date based (“Gemini Pro 2025”), or number + meaningful letter (eg 4o - “Omni”) for model names.

I would consider this a case of "expectation management"-based versioning. This is a release designed to keep Gemini in the news cycle, but it isn't a significant enough improvement to justify calling it Gemini 3.0.
I think it's reasonable. The development process is just not really comparable to other software engineering: It's fairly clear that currently nobody really has a good grasp on what a model will be while they are being trained. But they do have expectations. So you do the training, and then you assign the increment to align the two.
I figured you don't update the major unless you significantly change the... algorithm, for lack of a better word. At least I assume something major changed between how they trained ChatGPT 3 vs GPT 4, other than amount of data. But maybe I'm wrong.
The number is purely for marketing.

If you could get much better performance without changing the algorithm (eg just by scaling), you'd still bump the number.

Funnily enough, from early indications (user feedback) this new model would've been worthy of the 3.0 moniker, despite what the benchmarks say.
Agreed, can't everyone just use semantic versioning, with 0.1 increments for regular updates?
Regarding semantic versioning: what would constitute a breaking change?

I think it makes sense to increase the major / minor numbers based on the importance of the release, but this is not semver.

As I see it, if it uses a similar training approach and is expected to be better in every regard, then it's a minor release. Whereas when they have a new approach and where there might be some tradeoffs (e.g. longer runtime), it should be a major change. Or if it is very significantly different, then it should be considered an entirely differently named model.
Or drop the pretext of version numbers entirely since they're meaningless here and go back to classics like Gemini Experience, Gemini: Millennium Edition or Gemini New Technology
Would be confusing for non-tech people once you did x.9 -> x.10
What would a major version bump look like for an llm?
Going from English to Chinese, I guess? Because that would not be a compatible version for most previous users.
At least for OpenAI, a .5 increment indicates a 10x increase in training compute. This so far seems to track for 3.5, 4, 4.5.
I think it's because of the big jump in coding benchmarks. 74% on aider is just much, much better than before and worthy of a .5 upgrade.
The elo jump and big benchmark gains could be justification
> This will mark the first experimental model with higher rate limits + billing. Excited for this to land and for folks to really put the model through the paces!

From https://x.com/OfficialLoganK/status/1904583353954882046

The low rate-limit really hampered my usage of 2.0 Pro and the like. Interesting to see how this plays out.

Any word on what that pricing is? I can't seem to find it
Traditionally at Google experimental models are 100% free to use on https://aistudio.google.com (this is also where you can see the pricing) with a quite generous rate limit.

This time, the Googler says: “good news! you will be charged for experimental models, though for now it’s still free”

Right but the tweet I was responding to says: "This will mark the first experimental model with higher rate limits + billing. Excited for this to land and for folks to really put the model through the paces!"

I assumed that meant there was a paid version with a higher rate limit coming out today

I expect this might be pricier. Hoping not unusable level expensive.
The Long Context benchmark numbers seem super impressive. 91% vs 49% for GPT 4.5 at 128k context length.
Google has the upperhand here because they are not dependent on nvidia for hardware. They make and uses their own AI accelerators.
Keen to hear more about this benchmark. Is it representative of chat-to-document style usecases with big docs?
Looks like it's this benchmark [1]. It's certainly less artificial than most long context benchmarks (that are basically just a big lookup table) but probably not as representative as Fiction.LiveBench [2], which asks specific questions about works of fanfiction (which are typically excluded from training sets because they are basically porn).

[1] https://arxiv.org/pdf/2409.12640

[2] https://fiction.live/stories/Fiction-liveBench-Feb-20-2025/o...

"MRCR (multi-round coreference resolution)" for those looking for the link to Michaelangelo
Update: Gemini 2.5 also crushes fiction.livebench
I'm most impressed by the improvement on Aider Polyglot; I wasn't expecting it to get saturated so quickly.

I'll be looking to see whether Google would be able to use this model (or an adapted version) to tackle ARC-AGI 2.

I wish they’d mention pricing - it’s hard to seriously benchmark models when you have no idea what putting it in production would actually cost.
These announcements have started to look like a template.

- Our state-of-the-art model.

- Benchmarks comparing to X,Y,Z.

- "Better" reasoning.

It might be an excellent model, but reading the exact text repeatedly is taking the excitement away.

I’m sure the AI helps write the announcements.
Maybe they just asked Gemini 2.5 to write the announcement.
And it was trained on the previous announcements.
... which were also written by earlier Gemini versions.
LLMs all the way down
Not all the way. At the bottom are a bunch of unpaid writers and artists and a horde of low-paid mturk workers in Nigeria.
That's ok. AI will kill those off soon enough, and like all winners, rewrite history enough so that that inconvenient theft never happened anyway. It's manifest destiny, or something.
I love this comment. It made me laugh.

    > mturk workers in Nigeria
Serious question: Has anyone tested how much money you can actually make doing a month of Amazon Mechanical Turk? (It would make for an interesting YouTube video!) I am curious if it is middle class wages in very poor countries (like Nigeria). Some light Googling tells me that middle class salary in Nigeria is about 6K USD, so about 3 USD/hour (assuming: 50 weeks/year * 40 hours/week = 2000 hours/year). Is this possible with MTurk?
which was written by ChatGPT3.5
Well hey, OpenAI did the exact opposite, and nobody liked that either.
How did you measure “nobody” liked OpenAI announcements?
I think people were fine with OpenAI demos. They were less fine with not actually ever releasing the demoed tech.
I liked their announcements and demos and continue to like them.
To clarify, by "doing the opposite" I mean OpenAI releasing GPT-4.5, a non-reasoning model that does worse on benchmarks (but supposed to be qualitatively better). People shit on OpenAI hard for doing that.
> It might be an excellent model, but reading the exact text repeatedly is taking the excitement away.

This is the commodification of models. There is nothing special about the new models but they perform better on the benchmarks.

They are all interchangeable. This is great for users as it adds to price pressure.

Man, I hope those benchmarks actually measure something.
I would say they are a fairly good measure of how well the model has integrated information from pretraining.

They are not so good at measuring reasoning, out-of-domain performance, or creativity.

Sooner or later someone is going to find "secret sauce" that provides a step-up in capability, and it will be closely guarded by whoever finds it.

As big players look to start monetizing, they are going to desperately be searching for moats.

Reasoning was supposed to be that for "Open" AI, that's why they go to such lengths to hide the reasoning output. Look how that turned out.

Right now, in my opinion, OpenAI has actually a useful deep research feature which I've found nobody else matches. But there is no moat to be seen there.

If you've seen DeepSeek R1's <think> output, you'll understand why OpenAI hides their own. It can be pretty "unsafe" relative to their squeaky-clean public image.
They don’t hide reasoning output anymore?
I was looking at this the other day. I'm pretty sure OpenAI run the internal reasoning into a model that purges the reasoning and makes it worse to train other models from.

I might be mistaken, but originally the reasoning was fully hidden? Or maybe it was just far more aggressively purged. I agree that today the reasoning output seems higher quality then originally.

Sooner or later someone is going to find the "secret sauce" that allows building a stepladder tall enough to reach the moon.

It's called the "first step fallacy", and AI hype believers continue to fall for it.

I take this as a good thing, because they're beating each other every few weeks and using benchmarks as evidence.

If these companies start failing to beat the competition, then we should prepare ourselves for very creative writing in the announcements.

The improvements have been marginal at best. I wouldn't call that beating.
Reminds me of how nobody is too excited about flagship mobile launches anymore. Most flagships for sometime now are just incremental updates over previous gen and only marginally better. Couple that with the chinese OEMs launching better or good enough devices at a lower price point, new launches from established players are not noteworthy anymore.

It's interesting how the recent AI announcements are following the same trend over a smaller timeframe.

Phones are limited by hardware manufacturing, plus maybe the annual shopping cycle peaking at Christmas. People won't have bought multiple iPhones even in its heyday.

These LLM models were supposedly limited by the training run, but these point-version models are mostly post-training driven, which seems to be taking less time.

If models were tied to a specific hardware (say, a "AI PC" or whatever) the cycle would get slower and we'll get a slower summer which I'm secretly wishing.

I think the greatest issue with buying a new phone today is ironically the seamless migration.

once you get all your apps, wallpaper, shortcut order and same OS, you really quickly get the feeling you spent 1000$ for the exact same thing

100% agree with you.

But it needs to be seamless to remove any friction from the purchase, but at the same time if it feels the same then we felt like we wasted money.

So what I usually do is buy a different colored phone and change the wallpaper.

My MacBook was the same. Seamless transition and 2 hours later I was used to the new m4 speeds.

I wish I wish I wish Google put better marketing into these releases. I've moved entire workflows to Gemini because it's just _way_ better than what openai has to offer, especially for the money.

Also, I think google's winning the race on actually integrating the AI to do useful things. The agent demo from OpenAI is interesting, but frankly, I don't care to watch the machine use my computer. A real virtual assistant can browse the web headless and pick flights or food for me. That's the real workflow unlock, IMO.

    > I've moved entire workflows to Gemini because it's just _way_ better than what openai has to offer, especially for the money.
This is useful feedback. I'm not here to shill for OpenAI, nor Google/Gemini, but can you share a concrete example? It would be interesting to hear more about your use case. More abstractly: Do you think these "moved entire workflows" offset a full worker, or X% of a full worker? I am curious to see how and when we will see low-end/junior knowledge workers displaced by solid LLMs. Listening to the Oxide and Friends podcast, I learned that they make pretty regular use of LLMs to create graphs using GNU plot. To paraphrase, they said "it is like have a good intern".
> can you share a concrete example?

Upload a complicated PDF of presentation and ask for insights that require some critical thinking about them.

> Do you think these "moved entire workflows" offset a full worker, or X% of a full worker

It can replace many junior analysts IMO.

Cancelled my account long time ago. Gemini models are like a McDonalds Croissant. You always give them an extra chance, but they always fall apart on your hands...
For me, the most exciting part is the improved long-context performance. A lot of enterprise/RAG applications rely on synthesizing a bunch of possibly relevant data. Let's just say it's clearly a bottleneck in current models and I would expect to see a meaningful % improvement in various internal applications if long-context reasoning is up. Gemini was already one of my favorite models for this usecase.

So, I think these results are very interesting, if you know what features specifically you are using.

But they score it on their own benchmark, on which coincidentally Gemini models always were the only good ones. In Nolima or Babilong we see that Gemini models still cant do long context.

Excited to see if it works this time.

We have incrementally improved 1% better then we were yesterday. Our competition is 1 day behind us now.
Like! No trolling: This could be a sarcastic comment written by an LLM!
Was going to comment the same thing, which has been bugging me off lately on all announcements that start with "our" followed by empty superlatives. Happy to not be alone on this!
Glaringly missing from the announcements: concrete use cases and products.

The Achilles heel of LLMs is the distinct lack of practical real-world applications. Yes, Google and Microsoft have been shoving the tech into everything they can fit, but that doesn't a product make.

ChatGPT has like 500M weekly active users, what are you on about?
"Well, Ed, there are 300 million weekly users of ChatGPT. That surely proves that this is a very real industry!" https://www.wheresyoured.at/longcon/
Is that article trying to argue that 500M people every week are visiting ChatGPT for the first (or second) time after reading about it in the news?

If I'm being incredibly generous I will concede that this could have been the case for the first few weeks when it was making headlines, but it clearly isn't true now.

It would be literally impossible to keep up these figures for as long as ChatGPT has without a ton of repeat users. There simply aren't enough people/devices.

(comment deleted)
I would say Adobe is doing an excellent job of commercialising image manipulation and generation using LLMs. When I see adverts for their new features, they seem genuinely useful for normie users who are trying to edit some family/holiday photos.
Why not snooze the news for a year and see what’s been invented when you get back. That’ll blow your mind properly. Because each of these incremental announcements contributes to a mind blowing rate of improvement.

The rate of announcements is a sign that models are increasing in ability at an amazing rate, and the content is broadly the same because they’re fungible commodities.

The latter, that models are fungible commodities, is what’s driving this explosion and leading to intense competition that benefits us all.

AI labs, it seems, use a template for system cards as well. OpenAI stands out because they showcase their employees using their tools for various use cases, which is refreshing.
> with Gemini 2.5, we've achieved a new level of performance by combining a significantly enhanced base model with improved post-training. Going forward, we’re building these thinking capabilities directly into all of our models, so they can handle more complex problems and support even more capable, context-aware agents.

Been playing around with it and it feels intelligent and up to date. Plus is connected to the internet. A reasoning model by default when it needs to.

I hope they enable support for the recently released canvas mode for this model soon it will be a good match.

It is almost certainly the "nebula" model on LLMarena that has been generating buzz for the last few days. I didn't test coding but it's reasoning is very strong.
Is this model going to be restricted to paying users?
hi, here is our new AI model, it performs task A x% better than our competitor 1, task B y% better than our competitor 2 seems to be the new hot AI template in town
Can anyone share what they're doing with reasoning models? They seem to only make a difference with novel programming problems, like Advent of Code. So this model will help solve slightly harder advent of codes.

By extension it should also be slightly more helpful for research, R&D?

We're using it to RCA infrastructure incidents.
Seriously? That doesn't require a human?! Are we talking about some kind of "generic" incident? (Type 3: forgot to manually update the xxxx file.) Or what's going on?
(comment deleted)
Sounds unbelievable to me, but hey... :)

If theyre that easy, why not fix the casues for the needs for RCA? Our RCAs will not be solved by AI for decades, let me tell you that.

I found reasoning models are much more faithful at text related tasks too (i.e. 1. translating long key-value pairs (i.e. Localizable.strings), 2. long transcript fixing and verification; 3. look at csv / tabular data and fix) probably due to the reflection mechanism built into these reasoning models. Using prompts such as "check your output to make sure it covers everything in the input" letting the model to double-check its work, avoiding more manual checks on my end.
Have been using them for non-interactive coding where latency is not an issue. Specifically, turning a set of many free-text requirements into SQL statements, so that later when an item's data is entered into the system, we can efficiently find which requirements it meets. The reasoning models' output quality is much better than the non-reasoning models like 3.5 Sonnet, it's not a subtle difference.
Google has this habit of 'releasing' without releasing AI models. This looks to be the same?

I don't see it on the API price list:

https://ai.google.dev/gemini-api/docs/pricing

I can imagine that it's not so interesting to most of us until we can try it with cursor.

I look forward to doing so when it's out. That Aider bench mixed with the speed and a long context window that their other models are known for could be a great mix. But we'll have to wait and see.

More generally, it woud be nice for these kinds of releases to also add speed and context window as a separate benchmark. Or somehow include it in the score. A model that is 90% as good as the best but 10x faster is quite a bit more useful.

These might be hard to mix to an overall score but they're critical for understanding usefulness.

It's available now as an option in Google AI Studio and Google Gemini.
Thanks. I think my post lacked clarity of what I was talking about. I meant that most people care about API access to use with their favorite editor. It's a big limiter with grok, for example.

But I did mingle that with my knowledge of google's history of releasing without releasing these models which, as you point out, isn't true with this release.

It's "experimental", which means that it is not fully released. In particular, the "experimental" tag means that it is subject to a different privacy policy and that they reserve the right to train on your prompts.

2.0 Pro is also still "experimental" so I agree with GP that it's pretty odd that they are "releasing" the next version despite never having gotten to fully releasing the previous version.

(comment deleted)
How does Gemini have such a big context window?

I thought memory requirement grows exponentially with context size?

Memory grows linearly, compute grows quadratically (but with small constant - until ~100k the inference will be still dominated by non-quadratic factors).
Also reusing key/values for different queries can compress the KV cache, it can be an 1000x or 10000x improvement in bandwidth if the model is trained for it.
Just to clarify: simple prefix KV cache doesn't require any special model training. It does require the inference framework to support it, but most do by now.

You can see dramatic improvements in latency and throughput if there is a large shared prefix of the queries.

The Gemini 1.5 tech report do reference some papers about supporting large context window.
Funnyish story: the other night I asked my Pixel 9 to generate an image via Gemini, then I asked it to make a change. It didn't consider the previous context, so I asked it "Are you capable of keeping context?" No matter how clearly I enunciated "context", it always interpreted what I was saying as "contacts." After the 4th try, I said "context, spelled "c-o-n-t-e-x-t" and it replied with "Ah, you meant context! Yes..."

This stuff has a long way to go.

I noticed Gemini Flash 2.0 making a lot of phonetic typos like that, yeah. Like instead of Basal Ganglia it said Basil Ganglia.

I've also had it switch languages in the middle of output... like one word in the middle of a sentence was randomly output in some strange hieroglyphs, but when I translated them, it was the right word and the sentence made sense.

I was using the conversational feature of Gemini on my phone the other night and was trying to get it to read a blog post to me. The AI proceeded to tell me (out loud, via voice mode/speech synthesis) that it was a text based model and couldn't read text out loud.

For as amazing as these things are, AGI they are not.

In its defense: it probably is just a text model that hasn't been told that its output is being read to the user.
I think google is digging a hole for themselves by making their lightweight models be the most used model. Regardless of what their heavy weight models can do, people will naturally associate them with their search model or assistant model.
That might be considered fine if Google's larger goal is to make money from enterprises/Workspace integration, using consumer launches as splashy PR.

This way they get two rounds of headlines. "Gemini 2.5 released" and later on "Gemini 2.5 coming to all Google accounts."

Their willingness to integrate depends on their perception of the model quality.
TPUs have a network topology better suited for long context than gpus: https://jax-ml.github.io/scaling-book/tpus/#tpu-networking

> This nearest-neighbor connectivity is a key difference between TPUs and GPUs. GPUs connect up to 256 H100s in an all-to-all configuration (called a node), rather than using local connections. On the one hand, that means GPUs can send arbitrary data within a node in a single low-latency hop. On the other hand, TPUs are dramatically cheaper and simpler to wire together, and can scale to much larger topologies because the number of links per device is constant.

From the 2.0 line, the Gemini models have been far better at Engineering type questions (fluids etc) than GPT, Claude especially with questions that have Images that require more than just grabbing text. This is even better.
Google is overly cautious with their guardrails.

Granted, Gemini answers it now, however, this one left me shaking my head.

https://cdn.horizon.pics/PzkqfxGLqU.jpg

For better or worse, Google gets more bad press when their models get things wrong compared to smaller AI labs.
Ha, I still remember that super hilarious "You are under 18, so you should not write C++, as it is unsafe..." log from ... a year ago?
Looks like they're gradually removing guardrails, it returns Nixon for me.
Weird, they released Gemini 2.5 but I still can't use 2.0 pro with a reasonable rate limit (5 RPM currently).
Is Gemini and Bard same? I asked it a question and it said "... areas where I, as Bard, have..."
There is no point in asking such questions, the model doesn't know what it is on its own, and you could get many different answers if you repeat it a few more times.
Are Gemini and Bard same? I asked it a question and it said "... areas where I, as Bard, have...."
Normal Google rollout process: Bard is deprecated, Gemini is not ready yet.
I was recently trying to replicate ClaudePlaysPokemon (which uses Claude 3.7) using Gemini 2.0 Flash Thinking, but it was seemingly getting confused and hallucinating significantly more than Claude, making it unviable (although some of that might be caused by my different setup). I wonder if this new model will do better. But I can't easily test it: for now, even paid users are apparently limited to 50 requests per day [1], which is not really enough when every step in the game is a request. Maybe I'll try it anyway, but really I need to wait for them to "introduce pricing in the coming weeks".

Edit: I did try it anyway and so far the new model is having similar hallucinations. I really need to test my code with Claude 3.7 as a control, to see if it approach the real ClaudePlaysPokemon's semi-competence.

Edit 2: Here's the log if anyone is curious. For some reason it's letting me make more requests than the stated rate limit. Note how at 11:27:11 it hallucinates on-screen text, and earlier it thinks some random offscreen tile is the stairs. Yes, I'm sure this is the right model: gemini-2.5-pro-exp-03-25.

https://a.qoid.us/20250325/

[1] https://ai.google.dev/gemini-api/docs/rate-limits#tier-1

Update: I tried a different version of the prompt and it's doing really well! Well, so far it's gotten out of its house and into Professor Oak's lab, which is not so impressive compared to ClaudePlaysPokemon, but it's a lot more than Gemini 2.0 was able to do with the same prompt.