And why won’t the frontier models continue to become better? The open models are getting better but so are the frontier models. The frontier models might remain in a constant race to remain ahead
I don't think it's a matter of "staying ahead"; the proprietary frontier models are better, but the trouble for them is that open weight models are good enough in increasingly many cases. This raises the floor on the frontier companies and cuts their total addressable market by commodifying the easier LLM tasks. This is really the central argument of tfa :)
The big question is if there are larger economic gains to be made from ever greater intelligence. It seems to me like that may be the case for only a select few hard problems, while the vast majority of tasks approach their economic ceiling asymptotically with intelligence.
You suggest that the market segment for frontier models is necessarily shrinking. But what if it's just a failure of imagination on the part of users ?
They are running out of novel, clean training data and compute. There is probably a limit to how much improvement can be squeezed out of LLMs. Recent improvements have been more about orchestration and "reasoning" loops (i.e. iteratively feeding context back through the model).
For base models they really must be running out of new training materials. It's more about size and architecture. But they still seem to be getting big strides out of improving the post training. They keep dropping point upgrades in under 8 weeks lately, which is an insane pace for product release. At some point it's going to slow down but we're not close yet imo.
I suspect due to model distillation, their need to stay on top for IPO value is more important than absolutely crushing their competition, who will simply harvest their output and distill it for training data on a ~3 month delay. Also they are probably hitting a wall on increase in intelligence vs training time.
I cancelled my ChatGPT account for the restrictions placed on my account, inappropriately flagging about 10% of my queries as unsafe (I was writing grants in immunology). I haven't looked back. I will do the same if with Claude if Anthropic doesn't reverse course soon. What could I use instead? I find Grok very powerful and useful. Also, Google's Gemini, while also have some of the same restrictions, were at least sensible and not blindly blocking my prompts. So Grok and Gemini may be my go to AI's going forward
I somewhat agree, in that most of these life science adjacent demos are essentially "find good drug targets for $DISEASE", which mostly overfit to existing, well-classified drugs and targets. The biggest gains IMO will be in improved connectors with autonomous lab platforms, better sharing and annotation of relevant data sets, and yes also improving the pathway to clinical trials.
At the moment, it feels like releases like this overcommit and overpromise on "PhD level reasoning", which I wouldn't say is the absolute bottleneck in clinical research.
I am assuming the proposal is to knockout the gene Lim1, which in other animals, creates a brainless phenotype. You won't be able to swap a brain into this headless body (assuming it can fully develop), but this approach could be used for medical research and potentially solve the problem of organ donors, assuming it is ethical
Also, just because Lem1 creates a headless mouse doesn't mean it will do the same in Humans. But I suppose that's what the primate testing will reveal
I noticed how ChatGPT got progressively worse at helping me with my research. I gave up on ChatGPT 5 and just switched Grok and Gemini. I couldn’t be happier that I switched.
It's amazing how different are the experiences different people have. To me every new version of chatgpt was an improvement and gemini is borderline unusable.
A lot of people still have a shallow understanding of how LLMs work. Each version of a model has different qualities than the last, each model is better or worse at some things than others, and each responds differently to different prompts, styles. Some smaller models perform better than larger ones. Sometimes you should use a system prompt, sometimes you shouldn't. Tuning settings for the model inference (temperature, top_p, penalties, etc) significantly influence the outcome. (https://www.promptingguide.ai/introduction/settings, https://platform.openai.com/docs/guides/optimizing-llm-accur...)
Most "big name" models' interfaces don't let you change settings, or not easily. Power users learn to use different interfaces and look up guides to tweak models to get better results. You don't have to just shrug your shoulders and switch models. OpenAI's power interface: https://platform.openai.com/playground Anthropic's power interface: https://platform.claude.com/ For self-hosted/platform-agnostic, OpenWebUI is great: https://openwebui.com/
Gemini has a great model, but it's a bad product. I feel much happier using ChatGPT because Gemini just seems so barebones and unpolished. It has this feeling of a tech demo.
Scientific research and proof-reading. Gemini is the laziest LLM I've used. Frequently he will lie that he searched for something and just make stuff up, basically never happens to me when I'm using gpt5.2.
The way I summed it up to a friend recently is that Gemini 3 is smarter but Grok 4 works harder. Very loose approximation, but roughly maps to my experience. Both are extremely useful (as is GPT-5.2), but I use them on different tasks and sometimes need to manage them a bit differently.
Maybe they messed something up in the official interface then. I've heard that the PDF processing capabilities are also significantly worse in Gemini UI compared to using it through the API or Google AI Studio.
Any coding task produces some trash, while I can prototype with ChatGPT quite a lot, sometimes delivering the entire app almost entirely vibe-coded. Gemini, it takes a few prompts for it to get me mad and just close the tab. I use only the free web versions, never agentic ‘mess with my files’ thing. Claude, is even better than that, but I keep it for serious tasks only, so good it is.
Gemini loves to ignore Gemini.md instructions from the first minutes, to replace half of the python script with "# other code...", or to try to delete files OUTSIDE of the project directory, then apologise profusely, and try it again.
Utterly unreliable. I get better results, faster, editing parts of the code with Claude in a web ui, lol.
Odd, I've found that Gemini will completely fabricate the content of specific DOIs despite being corrected and even it providing a link to a paper which shows it is off about the title and subject of a paper it will cite. This obviously concerns me about its effectiveness as a research aide.
The old chat gpt models scanning the nih pub med repositories with proper prompting (e.g. …backed by randomized control trial data) was an amazing health care tool. The stripped down cheaper versions today are junk and I’ve had to start relying on grok :-( I’m not convinced OpenAI can make this work
