Unfortunately, myself and my team will be cancelling our paid subscriptions until this change is either reverted, toggled, or otherwise changed. Gemini is simply useless to us now.
9 Likes
I have an update regarding this matter. Logan Kilpatrick, one of the lead devs behind Gemini, is starting to see and respond to people’s concerns about thew new CoT on Twitter. If you care about this and want him to consider reverting it, you need to DM him immediately (@OfficialLoganK). Or alternatively, send him an email (hey@logank.ai).
Tell him about this thread. Tell him why you need the raw thinking process back. Tell him what you dislike about the summarized thought process. Tell him if you’re canceling your subscription over this. Make your voices heard. Be nice, but be persistent!
10 Likes
My user experience says the exact opposite. It has helped me (and others) a lot. It had a practical impact
4 Likes
Just to add my two cents here as well, the raw thought process was useful for me. It was good to check different ways to tackle a problem I was having, there were a couple of times that it gave me ideas.
Most of the time, it was a good way to check if gemini “internally” understood what I was talking about. There were a couple of times where I easily spotted it giving a biased response because it misunderstood my prompt, so it had irrelevant train of thoughts.
And other times, it was just plain fun and/or interesting to read what it was thinking.
As it is right now, it’s just… Meh. The summarizations are so macro that they don’t really say anything useful.
9 Likes
it was still extremely useful either way.
edit: wait did you just create an account to post this video and leave?
2 Likes
Yes, it’s not even remotely relevant to the actual issues the community is discussing here. Mods should remove that promotional so we can focus on the real problem.
7 Likes
It’s been an issue for me as well: Debugging own prompt, debugging Gemini thought process, and simply having access to more information.
I have been using the free Gemini plan so far, but I actually would even pay for access to the thinking processes as before.
2 Likes
I’m genuinely surprised. While other users have already articulated the usefulness and importance of Chain of Thought (CoT) extensively, I’d like to add some different points:
- Gemini is a highly advanced chatbot and can present challenges for new users. Furthermore, the detailed ‘Thoughts’ (CoT) could both impress and engage new users, showcasing its capabilities.
- The computational power is still being expended on generating the full chain of reasoning anyway. Hiding it and only showing a summary seems like an inefficient use of resources and an illogical move.
- Can competing companies utilize Chain of Thought (CoT) traces? They likely can and do. But who said that’s a bad thing? Open collaboration and learning could actually foster innovation across the industry.
- Mere ‘politeness’ is not what’s expected or needed from a truly advanced chatbot; transparency and accuracy are far more valuable.
- Gemini’s advanced functionality is, in a way, a collective result, profoundly shaped by user interaction, prompt refinement, and the ability to debug its internal processes. It’s not just about algorithms and data in isolation.
- Transparency in a chatbot’s operation is paramount, especially for experimental models. Otherwise, ethical questions regarding user interaction can easily arise. Google, unlike some other entities, has historically emphasized transparency and responsible AI development. This move seems to contradict the principles of open experimentation.
- Hidden errors and truncated reasoning paths, which CoT helped expose, now pose significant risks. After all, no one desires a regression in functionality, especially in critical debugging scenarios.
- Thought summaries inherently impact chat context. Who guaranteed that the loss of reasoning structure within chat contexts would not negatively impact generation quality? Personally, I hypothesize this will lead to a reduction in the depth of generation, perhaps fewer nested levels of thought or complex structures. Is this a desirable outcome for anyone?
Furthermore, we are only at the very beginning of our journey with experimental models. The notion of AGI by 2027 seems more like a fantasy than a reliable roadmap to achieving true AGI. Recurrent neural networks? Synthetic data? Nonsense.
And what is AGI, anyway?
9 Likes
The changes to the change on thinking has completely neutered an entire developmental field I was trying to research into. Chain of Thoughts can be structured and formatted into a framework to direct the model’s own internal reasoning. I was using this in its system instructions. “Summerzation” is not in any way more beneficial then actually being able to see the breakdown and fine tune the process. It needs to at the very least be an optional toggable feature if not brought back entirely. The effect on Gemini I’ve seen is akin to a lobotomy that has stripped it off any useful emergent properties that CoT provided, which is, as I’ve been experimenting with: Frameworked chain-of-thoughts.
8 Likes
This is a good decision by Google. Blame China/Deepseek.
“Modified by moderator” it’s not even my video 
God forbid I post something informative but ok.
2 Likes
Good decision for Google, not for developers.
While i can understand why Google did this, it still hurts the developers and understanding the model’s reasoning process.
Summaries are nowhere close to actual raw data.
1 Like
The lack of thinking is really making every new prompt to be less useful. Now the AI just replies, without considering nuances.
I’ve seen it missing many sarcastic moments and jokes that old models used to get.
2 Likes
I work on LLM-based applications that are in production and quickly scaling up, seeing a lot of demand, featuring difficult tasks. Our go-to model when implementing new hard tasks that benefit from reasoning used to be Sonnet 3.7 with thinking. After Gemini 2.5 Pro-preview was released in March, that became our new standard, as its all-round performance showed superior. We will now be switching back to Sonnet 3.7, because iterating on a prompt without being able to see the reasoning is like developing with a blindfold on.
“Why is the generation taking 3 minutes? Who knows, it keeps on thinking. Why is it making mistake A? Good question.”. This is not feasible for serious production usage.
The number of improvements we are able to make to our prompts and bugs we are able to catch by looking at the reasoning is (for Gemini, was) enormous.
The only ways we’ll be using Gemini 2.5 going forward is setting the thinking budget to 0 and only using it for 1. non-reasoning tasks or 2. when we’ve established that Sonnet 3.7 isn’t capable of the task in any reasonable manner, by doing old-school manual CoT (i.e. through a manual prompt-guided thinking step).
10 Likes
Very huge mistake, makes it harder to iterate or see where the model goes wrong of it’s just summary. It also makes learning harder for some tasks where you can see why the model came to a decision amd learn from it.
5 Likes
Yes also need to see the thoughts personally for research math occasionally.
2 Likes
The thinking trace is very useful because it:
- Tells you pretty fast if the model understood the assignment (if not, you can stop the response and edit your query right away)
- It is massively helpful in making you realize your prompting blind spots. The model understands the world massively differently than humans, and about a million things that are obvious to humans are not obvious to the model.
- During internal reasoning inference, it provides you with info/entertainment if you are interested. If not, no harm no foul. That’s why it’s an expandable box which is collapsed by default for all the most popular interfaces and includes “summaries” which I suppose are now summaries of the summaries inside AI Studio.
- Gemini 2.5 Pro tends to think for a long time, so this is a big factor. I suspect this results in technical people interested in the thinking trace being more likely to task switch and kind of forget about it sometimes (though this might be good for Google, I guess, in a way).
- The summaries are horrible. Removing them and just giving us the summaries of the summaries would be preferable, probably, in my opinion. This would just make running AI Studio lighter since it doesn’t run efficiently in the first place, and the thinking box introduces scroll glitches anyway, as of the last 3 months.
Right now it’s changed to a similar vibe to o3. It’s just doing things, but very frequently it’s completely misunderstanding the user’s intent, and the answer is more alien as a result (in my opinion)
3 Likes
Thank you to everyone who has shared their thoughts and concerns in this thread. We hear you. While we’re excited to now return thought summaries directly through the Gemini API for the first time, we understand this is a different experience from the raw thoughts previously available in AI Studio. It’s clear that in their current state, the summaries aren’t providing the depth and nuance you need, especially for debugging prompts and validating Gemini’s assumptions.
Please keep telling us what you think and how we can improve the summaries, or what specific aspects of the raw thoughts were most useful to you - it’s important for helping us get this right.
1 Like
Realistically, no amount of summarization engineering is ever going to be able to do the necessary job for real developer work. Raw thoughts have to be passed to the user to allow for real-world debugging. Any form of summary is going to lose information. Subtle changes in prompt language that will be lost in a summarizing LLM can be absolutely vital to actually figure out what’s wrong with a given prompt.
I feel the people in this thread have already very effectively described why raw thoughts are absolutely necessary, and why summaries are almost worse than just going non-thinking outright. There’s also a human factors element of feeling deeply patronized by being provided with a summary instead of the real, raw thought. We’re developers and power users. We can understand the raw thought process.
Thank you for actually listening. I think this is a relatively easy fix - throw the raw thoughts back in there.