Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Gemini 2.5 Deep Think

blog.google

Post reply on HN
Print view View original

91–100 of 259 posts

Re: Gemini 2.5 Deep Think

#91

Post by joaogui1 »

dweekly wrote: post #83
highfrequency wrote: post #3

Approach is analogous to Grok 4 Heavy: use multiple "reasoning" agents in parallel and then compare answers before coming back with a single response, taking ~30 minutes. Great results, though it would be more fair for the benchmark comparisons to be against Grok 4 Heavy rather than Grok 4 (the fast, single-agent model).

Dumb (?) question but how is Google's approach here different than Mixture of Experts? Where instead of training different experts to have different model weights you just count on temperature to provide diversity of thought. How much benefit is there in getting the diversity of thought in different runs of the same model versus running a consortium of different model weights and architectures? Is there a paper contr…

Mixture of Experts isn't using multiple models with different specialties, it's more like a sparsity technique, where you massively increase the number of parameters and use only a subset of the weights in each forward pass.

Re: Gemini 2.5 Deep Think

#92

Post by SubiculumCode »

irthomasthomas wrote: post #40

You can spin up a version of this at home using simonw's LLM cli with the llm-consortium plugin. Bonus 1: Use any combination of models. Mix n match models from any lab. Bonus 2: Serve your custom consortium on a local API from a single command using the llm-model-gateway plugin and use it in your apps and coding assistants. https://x.com/karpathy/status/1870692546969735361 > uv tool install llm llm install llm-conso…

1. Why do you say this is a version of Gemini deep think? It seems like there could be multiple ways to build a multiagent model to explore a space. 2. The covariance between models leads to correlated errors, lowering the individual effectiveness of each contributing model. It would seem to me that you'd want to find a set of model architectures/prompt_congigs that minimizes covariance while maintaining individual accuracy, on a benchmark set of problems that have multiple provable solutions (i.e. not one path to a solution that is objectively correct).

Re: Gemini 2.5 Deep Think

#93

Post by twobitshifter »

andsoitis wrote: post #38
foundry27 wrote: post #34

I started doing some experimentation with this new Deep Think agent, and after five prompts I reached my daily usage limit. For $250 USD/mo that’s what you’ll be getting folks. It’s just bizarrely uncompetitive with o3-pro and Grok 4 Heavy. Anecdotally (from my experience) this was the one feature that enthusiasts in the AI community were interested in to justify the exorbitant price of Google’s Ultra subscription. I…

it turns out that AI at this level is very expensive to run (capex, energy). my bet is that AI itself won't figure out how to overcome these constraints and reach escape velocity.

our minds are incredibly energy efficient, that leads me to believe it is possible to figure out, but it might be a human rather than an AI that gives us something more akin to a biological solution.

Re: Gemini 2.5 Deep Think

#94

Post by dumbmrblah »

irthomasthomas wrote: post #40

You can spin up a version of this at home using simonw's LLM cli with the llm-consortium plugin. Bonus 1: Use any combination of models. Mix n match models from any lab. Bonus 2: Serve your custom consortium on a local API from a single command using the llm-model-gateway plugin and use it in your apps and coding assistants. https://x.com/karpathy/status/1870692546969735361 > uv tool install llm llm install llm-conso…

Thanks! Do you happen to know if there any OpenWebUI plugins similar to this?

Re: Gemini 2.5 Deep Think

#95

Post by ccozan »

dweekly wrote: post #78
amelius wrote: post #67

Earlier quoted context omitted.

It could be that your problem was too simple to justify the use of Deep Think. But yes, Google should have figured that out and used a less expensive mode of reasoning.

"I'm sorry but that wasn't a very interesting question you just asked. I'll spare you the credit and have a cheaper model answer that for you for free. Come back when you have something actually challenging."

Actually why not? Recognizing problem complexity as a fist step is really crucial for such expensive "experts". Humans do the same.

And a question to the knowledgeable: does a simple/stupid question cost more in terms of resources then a complex problem? in terms of power consumption.

Re: Gemini 2.5 Deep Think

#96

Post by twobitshifter »

judge123 wrote: post #48

I'm wondering if 'slow AI' like this is a temporary bridge, or a whole new category we need to get used to. Is the future really about having these specialized 'deep thinkers' alongside our fast, everyday models? Or is this just a clunky V1 until the main models get this powerful on their own in seconds?

we’re optimizing for quality over performance right now, at some point the pendulum will swing the other way, but there might be problems that require deep thinking just like we have a need for supercomputers to run jobs for days today.

Re: Gemini 2.5 Deep Think

#97

Post by riwsky »

dweekly wrote: post #78
amelius wrote: post #67

Earlier quoted context omitted.

It could be that your problem was too simple to justify the use of Deep Think. But yes, Google should have figured that out and used a less expensive mode of reasoning.

"I'm sorry but that wasn't a very interesting question you just asked. I'll spare you the credit and have a cheaper model answer that for you for free. Come back when you have something actually challenging."

“This meeting could’ve been an email”

Re: Gemini 2.5 Deep Think

#98

Post by mNovak »

nickandbro wrote: post #43

Ladies and Gentlemen, Here's Gemini Deep Think when prompted with: "Create a svg of a pelican riding on a bicycle" https://www.svgviewer.dev/s/5R5iTexQ Beat Simon Willison to it :)

Interestingly it seems to draw the bike's seat too (around line 34) which then gets covered by the pelican.

Re: Gemini 2.5 Deep Think

#99

Post by declan_roberts »

childintime wrote: post #87

This comes at a time where my experience with Gemini is lacking, it seems to get worse. It's not picking up on my intention, sometimes replies in the wrong language, etc. Either that or I am just transparent that it's a tool and its feelings are hurt. I've had to call it a moron several times, and it was funny when it started reprimanding me for my foul language once. But it was wrong. This behavior seems new. I coul…

I had a good experience with gemini-cli (which I think uses pro initially?). It's not very good, but it's very fast. So when it's wrong it's wrong very quickly and you can either solve it yourself or pivot your prompt. For a professional software engineer this actually works out OK.

Re: Gemini 2.5 Deep Think

#100

Post by danenania »

LeafItAlone wrote: post #90
foundry27 wrote: post #34

I started doing some experimentation with this new Deep Think agent, and after five prompts I reached my daily usage limit. For $250 USD/mo that’s what you’ll be getting folks. It’s just bizarrely uncompetitive with o3-pro and Grok 4 Heavy. Anecdotally (from my experience) this was the one feature that enthusiasts in the AI community were interested in to justify the exorbitant price of Google’s Ultra subscription. I…

> It’s just bizarrely uncompetitive with o3-pro and Grok 4 Heavy. In my experience Grok 4 and 4 Heavy have been crap. Who cares how many requests you get with it when the response is terrible. Worst LLM money I’ve spent this year and I’ve spent a lot.

It's interesting how multi-dimensional LLM capabilities have proven to be.

OpenAI reasoning models (o1-pro, o3, o3-pro) have been the strongest, in my experience, at harder problems, like finding race conditions in intricate concurrency code, yet they still lag behind even the initial sonnet 3.5 release for writing basic usable code.

The OpenAI models are kind of like CS grads who can solve complex math problems but can't write a decent React component without yadda-yadda-ing half of it, while the Anthropic models will crank out many files of decent, reasonably usable code while frequently missing subtleties and forgetting the bigger picture.

Post reply on HN