by vinhnx on 4/12/25, 3:58 AM with 819 comments
by thunderbird120 on 4/12/25, 6:13 AM
[1]https://blog.google/products/google-cloud/ironwood-tpu-age-o...
by levocardia on 4/12/25, 7:14 AM
But I will admit, Gemini Pro 2.5 is a legit good model. So, hats off for that.
by codelord on 4/12/25, 6:51 AM
by sva_ on 4/12/25, 9:06 PM
For a short while, Claude was the best thing since sliced cheese, then Deepseek was the shit, and now seemingly OpenAI really falls out of favor. It kinda feels to me like people cast their judgement too early (perhaps again in this case.) I guess these are the hypecycles...
Google is killing it right now, I agree. But the world might appear completely different in three months.
by gcanyon on 4/12/25, 12:35 PM
An LLM-based "adsense" could:
1. Maintain a list of sponsors looking to buy ads
2. Maintain a profile of users/ad targets
3. Monitor all inputs/outputs
4. Insert "recommendations" (ads) smoothly/imperceptibly in the course of normal conversation
"In order to increase hip flexibility, you might consider taking up yoga."
Was generated because it might lead to the question:
"What kind of yoga equipment could I use for that?"
Which could then lead to the output:
"You might want to get a yoga mat and foam blocks. I can describe some of the best moves for hips, or make some recommendations for foam blocks you need to do those moves?"
The above is ham-handed compared to what an LLM could do.
by remoquete on 4/12/25, 6:01 AM
And now that I'm on it, I don't think I'm going back. Google did it again.
by tkgally on 4/12/25, 8:21 AM
That matches my impression. For the past month or two, I have been running informal side-by-side tests of the Deep Research products from OpenAI, Perplexity, and Google. OpenAI was clearly winning—more complete and incisive, and no hallucinated sources that I noticed.
That changed a few days ago, when Google switched their Deep Research over to Gemini 2.5 Pro Experimental. While OpenAI’s and Perplexity’s reports are still pretty good, Google’s usually seem deeper, more complete, and more incisive.
My prompting technique, by the way, is to first explain to a regular model the problem I’m interested in and ask it to write a full prompt that can be given to a reasoning LLM that can search the web. I check the suggested prompt, make a change or two, and then feed it to the Deep Research models.
One thing I’ve been playing with is asking for reports that discuss and connect three disparate topics. Below are the reports that the three Deep Research models gave me just now on surrealism, Freudian dream theory, and AI image prompt engineering. Deciding which is best is left as an exercise to the reader.
OpenAI:
https://chatgpt.com/share/67fa21eb-18a4-8011-9a97-9f8b051ad3...
Google:
https://docs.google.com/document/d/10mF_qThVcoJ5ouPMW-xKg7Cy...
Perplexity:
https://www.perplexity.ai/search/subject-analytical-report-i...
by pzo on 4/12/25, 10:07 AM
1) is dirty cheap ($0.1M/$0.4M),
2) is multimodal (image and audio),
3) reliable rate limit (comparing to OSS ml ai providers),
4) fast (200 tokens/s).
5) if need realtime API they provide as well for more expensive price (audio-to-audio)
It's my go to model for using as an API for some apps/products. https://artificialanalysis.ai/models/gemini-2-0-flash/provid...
by antirez on 4/12/25, 6:05 AM
by ruuda on 4/12/25, 6:15 AM
by flexie on 4/12/25, 8:49 AM
Brands matter, and when regular people think AI, they think of OpenAI before they think Google, even if Google has more AI talents and scores better on tests.
And isn't it good? Who wants a world where the same handful of companies dominate all tech?
by porphyra on 4/12/25, 11:35 AM
> I'm just a language model, so I can't help you with that.
https://g.co/gemini/share/cb3afc3e7f78
Chatgpt 4o correctly identified the guy as Ratner and provided the relevant quotes.
by Lukman on 4/12/25, 6:42 AM
It would decide arbitrarily not to finish tasks and suggest that I do them. It made simple errors and failed to catch them.
by zmmmmm on 4/12/25, 9:38 AM
I had an interesting experience: I was taking out a loan for some solar panels and there were some complicated instructions across multiple emails. I asked Gemini to summarise exactly what I had to do. It looked through my emails and told me I had to go to the web site for local rebate scheme and apply there. It even emphasised that it was important that I do that. I scoffed at it because I thought my installer was going to do that and wrote it off. A few weeks later, guess what: the installer calls me because they can't see the rebate application in their portal and want me to go check that I applied for it (!). Sure enough, I missed the language in the email telling me to do that and had to do exactly what Gemini had said weeks ago.
I do think Google has a real shot here because they have such an integrated email and calendaring solution where everyone already assumes it's online, fully indexed etc.
by CSMastermind on 4/12/25, 9:45 AM
Benchmarks aside Gemini 2.5 Pro is a great model and now often produces the best code for me but it's not notably better than any of the other frontier models in my testing each of which tends to have their own strengths and weaknesses.
And Google's wrapper around Gemini is easily the most frustrating of any of the major AI companies. It's content guardrails are annoying and I just learned yesterday it won't let you upload json files for whatever reason (change the extension to txt without modifying the contents in any way and it works just fine).
by ozgune on 4/12/25, 8:32 AM
* The article compares Gemini 2.5 Pro Experimental to DeepSeek-R1 in accuracy benchmarks. Then, when the comparison becomes about cost, it compares Gemini 2.0 Flash to DeepSeek-R1.
* In throughput numbers, DeepSeek-R1 is quoted at 24 tok/s. There are half a dozen providers, who give you easily 100+ tok/s and at scale.
There's no doubt that Gemini 2.5 Pro Experimental is a state of the art model. I just think it's very hard to win on every AI front these days.
by giorgioz on 4/12/25, 7:06 AM
I just open Google Gemini Android app and asked to generate a JS script with Gemini 2 Flash and did the same with ChatGPT.
Gemini did not highlighted with colors the code. ChatGPT did highlighted with colors the code.
Colors in code are extremely useful to grok the code and have a nice DX.
I'm sure if I dig into Gemini's product I'll find dozens of UX/DX ways in which ChatGPT is better.
Google is still playing catch-up with LLM products. ChatGPT is still the one making the announcements and then Gemini doing the same UX/use case enhancements weeks/months later.
by Waterluvian on 4/12/25, 7:01 AM
Not that I think Demis is or is not trustworthy, but I think it’s a bit foolish to believe it would be allowed to matter.
by labrador on 4/12/25, 10:06 AM
by ww520 on 4/12/25, 9:08 AM
I asked it whether a language feature in Zig was available. It answered yes and proceeded to generate a whole working sample code. I compiled it and got an error. Reported the error and it said the error showed I typed it wrong and asked me to make sure it's typed correctly. Eh?! It's a copy-and-paste. I confirmed again it's wrong. It then said it must be my compiler version was too old. Nope, using the latest. It then said very convincingly that based on its extensively research into the language official documentation, official examples, and release notes, the feature must exist. I asked it to show me the reference materials it used to draw the conclusion. None of links it gave were valid. I told it they were wrong. It gave back another set of links and claimed it had checked the links to make sure they are alive. The links were alive but didn't contain any mention of the feature. I let it know again. It admitted couldn't find the mentioned feature. But it insisted the feature had been merged in a PR. The PR link it gave was unrelated. I let it know. It gave me another 3 PR's and said one mentioned something related so the feature must be in. At the point I gave up.
The issue was that it sounded very convincing and stated "facts" very confidently, with backings to documents and other resources even if they were wrong or irrelevant. Even when told it gave the wrong info, it would double down and made up some BS reference material to back up its claim.
by a1371 on 4/12/25, 9:34 AM
by paradite on 4/12/25, 8:50 AM
Those products show OpenAI was innovating and leading in RL at that stage around 2017 to 2019.
by gwd on 4/12/25, 8:22 AM
A large player with massive existing streams giving away a product in a new market to undercut new entrants? Looks an awful lot like abuse of monopoly position...
by p0w3n3d on 4/12/25, 9:49 AM
by glimshe on 4/12/25, 6:28 AM
by DisjointedHunt on 4/12/25, 10:00 AM
They've got immense potential, sure. But to say that they're winning is a bit far from reality. Right now, their Cloud AI offerings to the enterprise are technologically superior to anything else out there from AWS, but guess what? AWS seems to have significantly more %age sales growth in this space with their larger base compared to GCP with their smaller market share.
The same can be said across turn based chat and physical AI. OpenAI continues to be the growth leader in the consumer space and a collection of Claude + self hosted + Gemini now in the enterprise / API space.
They need to be measuring themselves on moving the needle in adoption now. I'd hate for such amazing progress to stall out in a niche.
by brap on 4/12/25, 11:43 AM
Think about it. Whatever you’re trying to do online, either Search, Chrome or Android are in the critical path of like 90%+ of people if not more.
Once AI is deeply baked into these products, which are more like the “operating system” of the internet, the race is basically over.
Not to mention that Google is already sitting on the largest money printer in history and they can do this all day.
by godjan on 4/12/25, 10:21 AM
but Gemini and Claude still suck much worse then ChatGPT models
by csmpltn on 4/12/25, 11:41 AM
So many LLM workloads require high quality search results (backed by efficient, relevant, complete and up-to-date indexes), and that’s Google’s muscle.
by HarHarVeryFunny on 4/12/25, 12:49 PM
In a commodity business cost is key, and Google with their N'th generation home grown TPUs and AI-optimized datacenters have a big advantage over anyone paying NVIDIA markups for accelerators or without this level of vertical integration.
by aunty_helen on 4/12/25, 1:37 PM
Currently my teams building 2-3 systems based on Gemini, but trying to walk a client through setting up a gcp account and provision the model for video is a horrible experience. Chat et al would break their own backs trying to give you an api key fast enough, not google. Here’s a comically bad process with several layers of permissions that nobody asked for.
The irony of using ChatGPT to walk through setting up Gemini for a client was a highlight for me this week.
by dtquad on 4/12/25, 9:22 AM
What will happen with Google's AI wing when Google inevitably gets split up in the next 4-8 years?
by raindropm on 4/19/25, 6:30 AM
OpenAI have Sam as 'presenter' that people can hear him talk enthusiastically, Grok have Elon, even though that not really help much. Gemini? just AI product from faceless mega corporate. People don't feel relate to that. The emotion is part of equation of why people use something, not just function.
Also doesn't help with ChatGPT name already cemented in AI history. Every Bob and Jane heard about ChatGPT even they're use it once, but Gemini...who?
by admiralrohan on 4/12/25, 7:30 AM
X data is private now which would give advantage it real-time scenarios. And Chinese have made it state-level priority.
by Arch-TK on 4/12/25, 12:10 PM
Writing a visualiser and basic scrambler isn't hard to stumble upon, there's endless training material and not much to screw up. Writing a working solver even if you train it on examples would be hard.
Very funny.
by aprilthird2021 on 4/12/25, 6:46 AM
Sure, hindsight is 20/20, and who knows if any of these products will be big money makers vs commodities, and they may still fail at the productization of these things. Sure.
But insofar as productization follows great technology, Google was always going to have the upper hand here. It took many years but they did finally start coming out ahead
by mekpro on 4/12/25, 9:46 AM
by wiradikusuma on 4/12/25, 8:30 AM
BUT more often than not, it stopped halfway (the code, so it's unusable). I'm not sure if it's the plugin that cannot handle the response, but it never happens with Claude.
by FilosofumRex on 4/13/25, 3:18 AM
In healthcare, engineering, construction, manufacturing, or aviation industires adoption is mostly on the admin side for low priority/busy work - virtually no doctors, pharmacists, nurses, engineers, technicians or even sales people use LLMs on the job. LLMs products have serious quality issues and are decades behind industrial databases, simulation and diagnostic tools.
On the other hand in academics, consulting, publishing, journalism, marketing, advertising, law and insurance industries it's wildly adopted and is surging. For example, Westlaw's Co-counsel is better at drafting and case law briefing than all but the most experienced litigators. Not only it has not replaced lawyers, but is causing a boom in hiring since the time and cost of training new associates is drastically reduced and allows firms to take on more clients with more sophisticated case work.
by torginus on 4/12/25, 6:23 AM
by Alifatisk on 4/12/25, 10:39 AM
I've been using Qwen Chat a lot for the last couple of months because I got tired of Claudes small quota for free users, ChatGPTs inferior models and absurd pricing and Geminis (the previous models) heavy guardrails and censorship, like to the point that sincerely prompts actually triggers refusal.
I'll try Gemini 2.5 Pro Exp again and see how well it performs this time.
Also, did anyone notice that the ui of Google ai studio has changed? Can't find any mentions of this update in the release notes https://ai.google.dev/gemini-api/docs/changelog
by mmmBacon on 4/12/25, 11:40 AM
Whatever model is at the top can be surpassed if a competitor has enough compute scale. We are rapidly approaching the era where it’s difficult to have enough power in one campus. Distributed sites are needed if models continue to scale at 4.7x/year (see Epoch.ai) simply from a power perspective. You have to put the data centers where the power is and connect them together.
I believe the era of distributed training is already here however not everyone will be able to distribute training to multiple sites using their scale up networks. Their scale out networks will not be ready. So it could be that we see models plateau until distributed training infra is available.
I see the infrastructure side of AI and based on HW build out; Google has been slow to build and is behind everywhere.
by hadlock on 4/13/25, 1:35 AM
Is this a feature? I feel like using Google's LLM products only serves to feed their Ad machine to sell me more ads. Every cloud based office suite offers AI functionality now. Unless I'm doing something really complex/dramatic I'm going to choose the LLM that isn't tied to a giant machine selling me ads over the one that does every time. Chat LLM products have pretty much effectively allowed me to divorce myself from the Google Ad Machine, now that I'm free I'm not walking back willingly.
by GaggiX on 4/12/25, 7:16 AM
by oezi on 4/12/25, 6:09 AM
The key question is if the can stop the decline in search or pivot their revenue streams to Gemini.
by ahmedhawas123 on 4/12/25, 8:52 PM
- The last time I checked (3-4 months ago) Gemini embedding models are probably the least reliable / contextually aware out there - A significant chunk of the market will want the ability to use locally hosted models / manage their own which Google currently has no play for - API documentation. Across the big managed models they are likely the least well documented model. - Allowing for more system vs. user prompts
by siliconc0w on 4/12/25, 3:18 PM
It's also still uncertain whether Google can turn Gemini into a successful product that either consumers or businesses want to use. They are famously bad at translating their technological advantage into good products - for example the way they shoehorn AI chat into search just makes both worse (imo).
I think OpenAI has the consumers and that'll make it easier to get business. Once they start eating into Google's lunch with AI booked flights and hotels...
by xpe on 4/13/25, 1:25 AM
I'm so done with articles that don't even try to talk about probability sensibly.
The article doesn't make a good case even for a watered-down version of the claim. Where is the logic?
Until the author puts forth his model of change and how/why Google is unassailably ahead, I'm not buying his hyperbole.
> When I put the Google + DeepMind picture together, I can only wonder why people, myself included, ever became so bullish on OpenAI or Anthropic or even Meta.
Yikes. Hindsight bias in full display.
by cryptozeus on 4/12/25, 9:03 AM
by uejfiweun on 4/12/25, 9:00 PM
by ReptileMan on 4/12/25, 8:59 AM
Probably sending people to spend money at your competition is not the surefire way to market dominance.
by jerrysmith6478 on 4/14/25, 7:53 AM
by ein0p on 4/12/25, 5:13 PM
by benmathes on 4/16/25, 10:54 PM
The crux is not whether the extremely rich incumbent (google) will have better infrastructure, but whether that is the field of competition that matters.
by jerrysmith6478 on 4/15/25, 6:08 AM
by jerrysmith6478 on 4/15/25, 6:13 AM
by jerrysmith6478 on 4/15/25, 5:43 AM
by jerrysmith6478 on 4/15/25, 5:39 AM
by _blk on 4/12/25, 3:33 PM
by qwertox on 4/12/25, 8:26 AM
They really need to fix this.
It gets to a point where on each submit Google Chrome pops up a "wait | close tab" dialog.
I then have to use AI Studio for the "big picture" in one tab and ChatGPT in the smaller subtasks which help with the big picture.
by jerrysmith6478 on 4/15/25, 5:56 AM
by jerrysmith6478 on 4/14/25, 12:15 PM
by jerrysmith6478 on 4/14/25, 12:13 PM
by jonplackett on 4/12/25, 12:27 PM
Even for coding I find GPT4o to be more concise and write more sensible things.
I get the one-shot ‘build me a flight simulator’ type thing is special to Gemini 2.5 - but who actually ever uses it that way?
I feel a bit old school for aging it, but I still prefer ChatGPT at this moment. Am I the only one?
by jerrysmith6478 on 4/14/25, 12:52 PM
by jerrysmith6478 on 4/14/25, 11:55 AM
by jerrysmith6478 on 4/14/25, 11:09 AM
by jerrysmith6478 on 4/14/25, 10:42 AM
by jerrysmith6478 on 4/14/25, 9:22 AM
by jerrysmith6478 on 4/14/25, 11:00 AM
by jerrysmith6478 on 4/14/25, 9:46 AM
by jerrysmith6478 on 4/14/25, 7:53 AM
by jerrysmith6478 on 4/14/25, 9:34 AM
by nikolayasdf123 on 4/12/25, 7:10 AM
by alimhaq on 4/12/25, 4:01 PM
by keepamovin on 4/12/25, 9:36 AM
by renewiltord on 4/12/25, 7:36 AM
by microtherion on 4/12/25, 1:19 PM
by rvba on 4/12/25, 8:24 AM
Also the search quality itself went downhill. There was a great article about that on HN some time ago.
by aitchnyu on 4/14/25, 6:26 AM
by karel-3d on 4/12/25, 11:27 AM
If I want to use OpenAI models, I download ChatGPT app.
What do I need to do to use Google's model? They have so mamy things called Gemini... I genuinely have no clue
by retskrad on 4/12/25, 6:39 AM
by coolvision on 4/13/25, 7:09 AM
by cnych on 4/12/25, 1:49 PM
by upmind on 4/13/25, 12:30 PM
by croes on 4/12/25, 2:43 PM
All make simple mistakes, all hallucinate, all are not reliable.
by k2xl on 4/12/25, 11:31 AM
by nabla9 on 4/12/25, 8:38 AM
1) AI research as science and
2) Productization and engineering that science into something to sell.
While Google DeepMind focused on things that won Hassabis and Jumper Nobel prize in Chemistry, OpenAI took transformers architecture (Google researchers invented), built the first big model, and engineered it into a product.
Google has the best researchers, and does most research. When they finally chose to jump into the business and pull Hassabis and others from doing more important stuff to moneymaking, obviously they win.
by GrumpyNl on 4/12/25, 11:27 AM
by ohgr on 4/12/25, 6:51 AM
by silexia on 4/12/25, 8:08 PM
by davidmurdoch on 4/12/25, 12:06 PM
by sMarsIntruder on 4/12/25, 10:08 AM
by gigatexal on 4/12/25, 8:44 PM
by twism on 4/12/25, 1:13 PM
by cryptozeus on 4/12/25, 7:28 PM
by jjallen on 4/12/25, 1:56 PM
by conartist6 on 4/13/25, 12:10 PM
by patwolf on 4/12/25, 12:10 PM
I know that even if they never inject ads directly in Gemini, they'll be using my prompts to target me.
by nigel_doug on 4/13/25, 1:06 PM
by dostick on 4/13/25, 10:21 AM
by throwaway519 on 4/12/25, 4:07 AM
by not_a_bot_4sho on 4/12/25, 3:07 PM
by nullbio on 4/12/25, 11:58 PM
by phemartin on 4/12/25, 3:31 PM
by Aeroi on 4/13/25, 10:07 PM
by pcdoodle on 4/12/25, 10:12 AM
by lemonish97 on 4/12/25, 3:48 PM
by hm-nah on 4/13/25, 3:32 AM
by IG_Semmelweiss on 4/13/25, 3:05 AM
Is it really true the 2.5 is actually good ?
by zkmon on 4/12/25, 6:31 PM
by sidcool on 4/12/25, 6:31 PM
by nipperkinfeet on 4/12/25, 5:04 PM
by lofaszvanitt on 4/12/25, 9:21 AM
by Giorgi on 4/12/25, 9:09 AM
by asadalt on 4/12/25, 1:29 PM
by calmworm on 4/12/25, 7:36 PM
by cryptoegorophy on 4/12/25, 3:01 PM
by p1dda on 4/12/25, 11:25 AM
by indigodaddy on 4/12/25, 9:47 PM
- you can't locally install or onprem Gemini right, so why does small make it better for edge applications, essentially because small means light and fast, so it will respond quicker and with less latency? Requests are still going out over the network to Google though right?
by glacier5674 on 4/12/25, 6:14 AM
> Fred Alex Ottman, a retired American professional wrestler, is known for his WWF personas "Tugboat" and "Typhoon". He also wrestled as "Big Steel Man" and "Big Bubba" before joining the WWF in 1989. Ottman wrestled for the WWF from 1989–1993, where he was a key ally of Hulk Hogan. He later wrestled in World Championship Wrestling as "The Shockmaster", a character known for raising his fist and making a "toot-toot" sound.
Which is obviously false. The "toot-toot" was part of his gimmick as Tugboat, while the Shockmaster gimmick is known for its notoriously botched reveal.
Point being, Google is losing on the "telling one early 90s wrestling gimmick from another" AI front.