The first commit was in 2023. It's been only ~3.5 years which feels like an eternity in the ( recent ) LLM world.
llama.cpp v1.0.0 is now available
huggingface.co/blog/ggml-org…
12
I am very curious about the impact of the price increase and hence the usage reduction of deepseek models in opencode-go? Is it something drastic or people are still continuing the usage.
Anything you can share @thdxr ?
1
50
Respect !
TLDR: I’ve started a small fund from my Base44 money that’s focused on cancer research and breakthrough therapies.
I’ve invested in 7 incredible companies in the past 4 months and plan to do more as my liquidity allows. Listing my investment thesis below
I welcome anyone like me who has been blessed financially to join me and find ways to finally manage or even end this disease.
Longer -
I honestly believe cancer is humanity's worst enemy.
~40% will be diagnosed with it, and out of which a ~third won’t make it. And numbers are currently not getting dramatically better.
For many it might sound like a statistic
But for some of us who have been there ourselves or had to go through it with a person we love, it’s one of the worst things you can imagine.
I had to go through it with my mom - who passed away 4 months ago and whom I miss and think about every single day since.
This field needs a dramatic push.
The current wave of technology and AI is doing a lot of good in the world,
But cancer therapy is one of those things that are not moving nearly fast enough.
And there’s no good reason for that - AI is creating so much opportunity right now.
Here's my thesis on how things will play out:
For drug development, the bottleneck will shift to
1. Manufacturing
2. Real world testing
"In silico" is going to accelerate incredibly fast. AI is identifying new target proteins and designing new molecules, drastically shortening timelines.
AI will create an abundance / inflation of ideas. It will help design molecules faster than anything historically possible.
This is already happening.
Play with Claude Science or Codex and you'll be surprised at how far you can get with just a few prompts.
We need a better way to *manufacture and test* the 1000s of ideas coming from AI models spitting out potential drug candidates
1. Manufacturing: i.e. how fast can we synthesize a potential drug
Not too far into the future, companies with manufacturing excellence will win over companies with the smartest scientists.
Being able to synthesize a drug candidate fast - together with efficient methods to test efficacy (more on that below), basically means you get more shots at a target hence more chances at succeeding.
This is the same exact lesson the software industry learned about the importance of fast iterations.
Companies like Starget Pharma are leveraging AI heavily to iterate on drug candidates + figured out in-house manufacturing in a way that lets them iterate on drugs almost as fast as if they were a software company iterating on features.
2. Real world testing: i.e. how fast can we get a call on a drug's efficacy.
We need better real-world indicators for drug efficacy than mouse / animal models.
The success rate of animal models -> human clinical trials is less than 8%.
Breaking down this bottleneck will change the entire industry.
Developing an oncology drug and taking it through all clinical trials usually cost >> $ 100m
Thus taking the wrong bet is devastating (but still, happens a lot).
Giving drug development companies the ability to better understand the efficacy and mechanisms of their drugs will change a lot. Esp in the age of abundant drug candidates.
Furthermore - and once those methods proved efficient - Joining forces with regulators to shorten (or even skip) some clinical trials altogether will cut years from the process and save many lives by bringing new drugs to market faster.
CuResponse is pioneering functional testing - taking real cancer biopsies and showing drug efficacy with very high precision. Can be used for precision oncology as well as testing new drugs.
Cellint is providing a fully automated cell-culture R&D platform for companies and researchers.
----
I’d love to see a future where there are “cloud services” for drug development companies.
A company should eventually be able to submit a drug candidate programmatically and receive:
1. The synthesized drug.
2. Testing across cells, tissues, organoids, and tumoroids.
3. Structured results showing what worked, what failed, and why.
4. A recommended next iteration derived from the biological reactions.
I’ve invested in all of the above companies and am searching for additional companies to bring this “cloud services for drug dev” vision to life.
----
Alongside it, I am also investing in direct therapeutic moonshots.
I invested in Baccine, which is developing a bacteria-based cancer immunotherapy platform and deserves an entire separate post.
I’m also a proud LP in Even One Ventures - @sytses ’s fund to fight cancer. Sid and Jacob Stern are an inspiration to me as I’m making my first steps in the field, and I bet their companies will make a massive impact.
And 3 other incredible companies focused on how therapies are delivered - where I’ll elaborate in future posts.
----
I’m still obviously spending most of my time on Base44. Which is a plus for the companies - as we’re able to support them with tools, software, AI agents, etc.
I’m looking for people to join me - invest side by side or help one of the companies.
I don’t have a name for the fund yet, nor an email or a website
I will get to it this weekend and will post it in the comments.
2
49
Indeed ! But curious to know why do you folks don't support prompt caching? Is there a technical decision/limitation behind this?
1
36
Oh wow !!
Most people should probably update their priors on the state of open-source speech-to-speech.
It's honestly kind of mind-blowing.
We teamed up with @cerebras to build a fully open-source realtime voice demo (models + code) to show what's possible today.
Demo : huggingface.co/spaces/smolag…
Blog: huggingface.co/blog/cerebras…
Go test it, fork it, tweak it, and impress your friends.
video is raw, no cut, no speed-up, first take
2
51
Learned the hard way that caching is not implemented on Nebius Token Factory @nebiustf
Deepseek v4 pro ~40M input few minutes cost 70$
What is the background behind this? (caching is in general good for both the model providers and for customers)
1
23
hunk is indeed great, but kinda surprised that not many folks mentioned "delta" github.com/dandavison/delta
Hunk is very good. It has completely replaced any other local diff viewer for me. It looks good, its speedy, good keyboard shortcuts, good mouse support for fallback. Great software @bentlegen. github.com/modem-dev/hunk
36
These folks are pushing the benchmark for open source, local hardware LLMs to another level
⚡ Meet Qwen3.6-35B-A3B:Now Open-Source!🚀🚀
A sparse MoE model, 35B total params, 3B active. Apache 2.0 license.
🔥 Agentic coding on par with models 10x its active size
📷 Strong multimodal perception and reasoning ability
🧠 Multimodal thinking + non-thinking modes
Efficient. Powerful. Versatile. Try it now👇
Blog:qwen.ai/blog?id=qwen3.6-35b-…
Qwen Studio:chat.qwen.ai
HuggingFace:huggingface.co/Qwen/Qwen3.6-…
ModelScope:modelscope.cn/models/Qwen/Qw…
API(‘Qwen3.6-Flash’ on Model Studio):Coming soon~ Stay tuned
21
Removing the auth file ~/.pi/agent/auth.json and logging in again fixed it for me.
Folks using pi, is anyone else facing this issue? It logs out after sometime of inactivity, noticed it with both github copilot and chatgpt logins.
Started noticing it today and stumbled upon this github issue - github.com/badlogic/pi-mono/…
cc: @badlogicgames
17
Folks using pi, is anyone else facing this issue? It logs out after sometime of inactivity, noticed it with both github copilot and chatgpt logins.
Started noticing it today and stumbled upon this github issue - github.com/badlogic/pi-mono/…
cc: @badlogicgames
Ankur Srivastava retweeted
blog.rust-lang.org/inside-ru…
test out async closures pls thx!!!
Ankur Srivastava retweeted
Our app gets hundreds requests per second, we've been dumping logs for a whole week and only used up, like, 6GB or something. Love it.
Also no messy Kibana/Grafana/Elastic setup or anything.
And I've just investigated a very suspicious incident by querying the log database 🩷
Ankur Srivastava retweeted
SaaS people, if you're looking for a log management app try out @OpenObserve
Not affiliated, no idea who they are, but the free-tier is generous AF (200GB per month, 15 days retention - enormous!)
PS. and if you're not dumping your nginx logs anywhere u're asking for trouble
Ankur Srivastava retweeted
I'm speaking tomorrow @fosdem in the #rustlang devroom:
How Pydantic V2 leverages Rust's Superpowers, Using Rust to build Python extensions
Would be great to meet @pydantic users, let me know if you're there.
fosdem.org/2023/schedule/tra…
Ankur Srivastava retweeted
The best coding tip I learned: "first make it work, then refactor and improve it."
Ankur Srivastava retweeted
💶📕 For those in Germany: Amazon.de is now selling my book for € 39.99:
amazon.de/-/en/Mara-Bos/dp/1…
It was a cool project 🙌. While deploying this on prod, we also made several improvements in the framework.
1
103




















