Dominik Lukes
@techczech
Followers
2K
Following
3K
Media
2K
Statuses
14K
Exploring applied epistemology, AI and metaphor. Current work on https://t.co/XxygzOGYN0.
UK
Joined April 2009
My periodic "big tech" appreciation post. I am very much an Open Source hippie at heart but big tech has made the world a better place for billions of people and has also provided so much of the infrastructure for the small and open tech, too. I have said nasty and derogatory.
0
0
3
There cannot be too many Open Source ASR models. Well done @MistralAI for keeping the engine of progress going here. I remember as recently as 3 years ago people saying that speech recognition was never going to be open because of the computational and data demands.
0
0
0
THIS WAS SUPPOSED TO BE A SHORT QUIP! BUT IT BECAME A TREATISE ON METHOD. OH WELL. This is the big epistemic dilemma of our time: aggregate statistic vs the individual case. You can learn more about the nature of things from a case but it's very easy to overlearn. You can learn.
I always learn a lot more from in-depth analysis of few random cases over dashboards of aggregate statistics across all cases. Both projections can be helpful but the latter is disproportionately pervasive.
0
0
0
The thing people get most wrong about metaphor is that it only explains the unfamiliar with the familiar. Often, the thing we are explaining is more familiar to us than the thing we are comparing it to and our perspective on the source changes more than the perspective on the.
@GregKamradt A model… but it’s much much more malleable and clearly has unique properties on lots of layers (like short vs long vs medium memory, etc). And it’s definitely a reasoning model. We emit thought tokens all the time. And clearly multimodal (even smell and taste, should we give.
0
0
1
More and more, LLM training recipes are starting to resemble designing a training program for humans. Not the actual learning process but the selection or creation and then the sequencing of how the training data is presented. Exactly what a syllabus designer will do.
Scaling up RL is all the rage right now, I had a chat with a friend about it yesterday. I'm fairly certain RL will continue to yield more intermediate gains, but I also don't expect it to be the full story. RL is basically "hey this happened to go well (/poorly), let me slightly.
0
0
2
My initial impressions of Kimi K2 by @Kimi_Moonshot. It took me a while to get to it, there seems to be a new open model from China every day but this one does 'feel' important - a definite frontier model smell:. - o3 feel on agentic choices when it comes to search.- Claude 4.
🚀 Hello, Kimi K2! Open-Source Agentic Model!.🔹 1T total / 32B active MoE model.🔹 SOTA on SWE Bench Verified, Tau2 & AceBench among open models.🔹Strong in coding and agentic tasks.🐤 Multimodal & thought-mode not supported for now. With Kimi K2, advanced agentic intelligence
0
1
3
Every generation builds a world for which they didn't have "the education" and then insists "the education" is essential to live in it.
0
0
2
How much weight should we give to calls for more critical thought when appended to statements that are themselves examples of the exact opposite?.
It’s no coincidence that the rise of AI correlates to the rise of fascism. Now more than ever, we can’t give up on critical thought.
0
0
1
Look, if you publish a book in 2025 that says 40% of GitHub Copilot written code contains security vulnerabilities and only mention in a footnote that's it's based on a study from 2022, you are pretty much just plain lying to your readers!.
0
0
1
I often think abou this: The Taiping rebellion was started by a guy who failed the Imperial examination for the fourth time and had a dream he was the brother of Jesus. It killed more people than WWI and is an example of the dual evils of high-stakes exams and religious imagery.
1
0
1
Common @AnthropicAI - this is just pathetic. You need to do better on sign in. Anything you can do to make this less user hostile @alexalbert__ ? Having to re-login into Claude is starting to fill me with dread.
0
0
0
I despise cookies and often wish bad things on those involved in inflicting this idiocy on the world while I'm trying to find where to make a badly designed popup go away wasting my life on navigating the petty rituals of ethics bureacracy.
2
0
1
Great study by @METR_Evals and I appreciate the thoroughness and lack of clickbaitiness. Yes, this is about limits of the usefullness of LLMs in some scenarios but I think the key lesson is one of context and knowledge of the context. This is the bit in the full paper that.
We ran a randomized controlled trial to see how much AI coding tools speed up experienced open-source developers. The results surprised us: Developers thought they were 20% faster with AI tools, but they were actually 19% slower when they had access to AI than when they didn't.
0
0
0
Yet another 'AI-first' browser. Having played around with Dia and Arc, I've been finding it hard to switch as a browser power-user but I can imagine people with less muscle memory might find this an improvement.
0
0
1
Ok, here's a surprisingly difficult puzzle: How do you definitively and with 100% certainty answer the question: "How many different women are mentioned in Pride and Prejudice?" with and without AI!.
0
0
1
RT @ankitdp_: You asked, and it's here: Gems with Canvas.Make your own custom experts with Gemini Gems and get them to write docs and apps….
0
27
0
There's potential to bite-sized learning except in practice the bite size displaces the learning. Yes! Learn in small bites, but take those bites often and chew them well. What this often devolves into is pretty little amuse bouche that satisfy the eye of the chef but feed noone.
1
0
1
LLMs have no mental checklist. Here are some slides with examples on what I mean by it. @AnnaRMills
1
0
4
Has anybody used a prompt framework recently or ever? I see them still share as if they're some big revelation of a secret. My prompting approach is - ask for what you want, see what you get and go from there. It's called a chatbot for a reason!.
2
1
3