Fine tune or point TxGemma, new AI model, on. all of our fav drug classes?

[yaesutom]

Has anyone else gone deep into AI land the last few years? I have been busy but I have gotten really fast at quickly “vibe coding” something I want into a working thing, so maybe I’ll get around to doing something with this soon. It’s open source and all of the AI stuff is super accessible these days.

https://developers.google.com/health-ai-developer-foundations/txgemma


One time a while back I dumped Pihkal into some LLMs and asked it to create some new psychedelics. It did a remix
of the book

The website says it can predict a bunch of stuff about molecules. It can be fine tuned, which is a way to make it better for your tasks. I have been wanting to get more into the training and fine tuning so maybe I’ll experiment. I want to ask it to make some new, novel stimulants.

 

[Rxcko]

Has anyone else gone deep into AI land the last few years? I have been busy but I have gotten really fast at quickly “vibe coding” something I want into a working thing, so maybe I’ll get around to doing something with this soon. It’s open source and all of the AI stuff is super accessible these days.

https://developers.google.com/health-ai-developer-foundations/txgemma


One time a while back I dumped Pihkal into some LLMs and asked it to create some new psychedelics. It did a remix
of the book

The website says it can predict a bunch of stuff about molecules. It can be fine tuned, which is a way to make it better for your tasks. I have been wanting to get more into the training and fine tuning so maybe I’ll experiment. I want to ask it to make some new, novel stimulants.

(For some reason I had to make a new account) Yeah I have been pushing the limits of these ai llm models. Deepseek & Google have been the ones staying true to keeping a science-first mentality. Grok has come around now. OpenAI loosened up a bit finally (maybe bc of my x1000 complaints to #freescience ). Claude can be a pain still, but I’ve always been able to bypass through discussing real world situations that I actually need the ai llm model for, so I’ll usually get the answers after a few tries.
Just recently, I found out that this is actually a technique ppl use, it’s called ‘echo chamber’ hacking. Had no clue but yeah pretty much the same.
Anyway, I actually had Claude do a whole project for specific promising novel stimulant(s). I had Manus AI put together a great presentation as well. Different topics but still in the same novel therapeutics / biotechnology space. The future is looking bright. Been in the chemistry biotech industry for a bit, and ai has opened up a whole new world for non coders.

 

[dopamimetic]

but I have gotten really fast at quickly “vibe coding” something I want into a working thing

I tried vibe coding with different AIs and they all suck. Claude was the best (before GPT-5, dunno about that one), but it does tons of silly mistakes when writing code. Like forgetting closing braces. Don't know whether I'd trust them with drugs. I think the current AI is more of a glimpse of what the future will become than ready yet. But of course for more simple stuff like multiple choice exams it's good.

 

[yaesutom]

I tried vibe coding with different AIs and they all suck. Claude was the best (before GPT-5, dunno about that one), but it does tons of silly mistakes when writing code. Like forgetting closing braces. Don't know whether I'd trust them with drugs. I think the current AI is more of a glimpse of what the future will become than ready yet. But of course for more simple stuff like multiple choice exams it's good.

There is actually a way around all these issues, check outCoding on a budget/using context tools, skipping the trendy agents for the thinking / difficult things
(you have to click the link on there, i think that is just the comment thread)

What is happening is people are trying to use agents for everything, and then needing to use the most expensive (Claude Sonnet and Opus) AIs - I noticed that the smarter the AI, the more likely it is to leave out a } - but its like me, i suck at writing 'proper english' as I am ADHD and for whatever reason hate it, but i'm not dumb. The answer to that is to feed the output into a dumber, more agentic AI like GPT 4.1 (highly underrated model, because it is like a person in the military - it does EXACTLY as you tell it, no more, no less). In an agentic environment they usually have just one model doing everything which takes away intelligence from whatever problems you are trying to solve. I know i'm gonna be right about this lol they got it all wrong. Should be, smart model for planning and problem solving and creation, brainstorming, etc, with ZERO tools and MCP stuff. Then, if you need to do something with the output, tell it to write a prompt for a dumber agent AI model, break it into smaller steps that a dumb agent can do, then you feed that into GPT 4.1. 4.1 will fix all the little glitches like missing )'s.. and you won't even notice it. Also dumber agents are better at doing the things like file editing, web searches, MCP tools, they just do what they're told and tend to output perfect English and code lol

 

[yaesutom]

(For some reason I had to make a new account) Yeah I have been pushing the limits of these ai llm models. Deepseek & Google have been the ones staying true to keeping a science-first mentality. Grok has come around now. OpenAI loosened up a bit finally (maybe bc of my x1000 complaints to #freescience ). Claude can be a pain still, but I’ve always been able to bypass through discussing real world situations that I actually need the ai llm model for, so I’ll usually get the answers after a few tries.
Just recently, I found out that this is actually a technique ppl use, it’s called ‘echo chamber’ hacking. Had no clue but yeah pretty much the same.
Anyway, I actually had Claude do a whole project for specific promising novel stimulant(s). I had Manus AI put together a great presentation as well. Different topics but still in the same novel therapeutics / biotechnology space. The future is looking bright. Been in the chemistry biotech industry for a bit, and ai has opened up a whole new world for non coders.

I actually am a coder but I wasn't really coding much because my brain just gets too bored after a certain amount of time and then I miss out on that needed dopamine hit for 'feedback' so I keep at something. With AI, the iteration loop is so much faster, and for people with various forms of ADHD or whatever other quirks, its great. I get a constant stream of daily dopamine hits fast enough to keep my brain happy and locked onto the tasks. I have been coding for like 14 hours a day.. when I wake up, and cannot even hardly force myself to leave the computer! Coding is so addictive now. This is the greatest thing to happen like, ever... for me. It brought back the passion like .. its so insane. I thought it was gone for good, I thought I was going to be just going thru life never enjoying anything again because of my weird brain. Thank god for AI

 

[Deleted member 576750]

I tried vibe coding with different AIs and they all suck. Claude was the best (before GPT-5, dunno about that one), but it does tons of silly mistakes when writing code. Like forgetting closing braces. Don't know whether I'd trust them with drugs. I think the current AI is more of a glimpse of what the future will become than ready yet. But of course for more simple stuff like multiple choice exams it's good.

gpt5 is phenomenal in my experience

I've had good experiences with previous openai models for coding but they required far more guidance and trial and error and lost context quickly. the introduction of "thinking" mode with multiple agents in o3 was a game changer and significantly improved code quality and reduced likelihood of weird issues

using these models for coding is different than we would code on our own of course, I've found that leaning on spec documents as a source of truth improves quality even further

agent mode with direct repo and terminal access could lead to fully autonomous programming