Karsten Konrad 🇪🇺 @zabong69.bsky.social · 4hI wouldn’t know how to spend this much on this, as I am using two agents max in parallel, and sometimes go to sleep. 050 5
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 4hThere is a new sheriff in town, and he works for peanuts. 022 2
Reposted by Karsten Konrad 🇪🇺Thomas Wood @advanced-eschatonics.com · 6hThe deepseek-v4.1 technical report is full of good references. I've been feeling burned out working on agents and clients lately. Spending today catching up with the field.arxiv.orgYou Only Index Once: Cross-Layer Sparse Attention with Shared RoutingLong-context inference in modern LLMs is increasingly constrained by decoding efficiency, especially in reasoning-heavy settings where models generate long intermediate chains of thought. Existing spa... 2407 40
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 4hThe way you paraphrased it clearly shows it can easily be misunderstood, the „it“ was actually meant as „making jokes about algebra“. As a German, I am new to zis humor thing. 010 1
Reposted by Karsten Konrad 🇪🇺Karsten Konrad 🇪🇺 @zabong69.bsky.social · 4hIt has been brought to my attention that my „normie“ comment needlessly belittles people outside of ML. I should clarify that I meant in the way that a lot of us in ML are nerds to the extreme, socially awkward types. I sincerely apologize for my choice of words. 041 4
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 4hIt has been brought to my attention that my „normie“ comment needlessly belittles people outside of ML. I should clarify that I meant in the way that a lot of us in ML are nerds to the extreme, socially awkward types. I sincerely apologize for my choice of words. 041 4
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 4hAgain, I can only apologize. You are right, my comment can be understood this way. 100 1
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 4hI think I should clarify that many of the people I met in the ML space are not neurotypical. I am not for instance. I will apologize, I did not mean “peasants“ but more people who wouldn’t make a joke about algebra. 100 1
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 5hIf you use mean for anything related to income or wealth, I will help Matt to slash your tires. 120 2
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 5hI could do this with cellular automata. Would hit even harder, Mandelbrot is at least math in complex plane. 030 3
Reposted by Karsten Konrad 🇪🇺Laurie Voss @seldo.com · 18hI'm going to keep ringing this alarm bell because we are leaving a whole cohort of new grads behind and it threatens to expand to a whole generation. We need to figure out how to train people to work in an AI first company and we have not figured out how to do that yet. 4534 53
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 6hTNG is underrated in how much influence it had. Now please make me that replicator. 170 7
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 13hThat „it’s all just linear algebra“ was an insider joke in ML circles, and we should have never let any normies hear it. They can’t handle it. 6411 41
Reposted by Karsten Konrad 🇪🇺Mark Riedl @markriedl.bsky.social · 16/10/2026The other thing that blows my mind is that I haven’t come across anyone talking about sim2real for LM agents. In robotics it is helpful to train in simulation where accidents don’t have consequences. Then fine-tune on reality. Why are we not doing this for LM agents? Don’t tell me it’s too expensive 142 4
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 16/10/2026This has been the most productive week of 30 years. 171 7
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 16/10/2026A lot is explained by Anthropic people not even getting their own rooms. 1220 22
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 16/10/2026I think this is true. The neural network is creating a fiction, and that fiction is conscious in the sense that a fictional character is conscious. What @colin-fraser.net has said, only the Sherlock example clicks more with me. 151 5
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 16/10/2026Funny, all my stuff now includes dashboards. Agents can move to review, but not on Done. Also, cards lets you confirm assumptions. 110 1
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 16/10/2026I have the theory that us speed readers will do just fine with these things. Exhausted, yes, but this feels like a superpower rn. 110 1
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 16/10/2026Very nice. Such a relaxing sound landscape. 000 0
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 16/10/2026Yes! 100% of success in agentic AI depends this: you have a clear KPI, and failure is both cheap and insightful. If you can somehow (re)define your task in this way, you will be gold. 010 1
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 16/10/2026This is the first time I see an actual experimental setup to answer a question around AI consciousness. This is how it should be done. 181 8
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 16/10/2026More likely that a consulting company sets a limit so that the consultants charge all additional use to the projects they are working in. 021 2
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 15/10/2026SQL is the essence of the human experience, if you believe them. 020 2
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 15/10/2026My recent experiments of combining Vega Lite with DuckDB is that BI software suites are dead men walking. All of them. I was able to replicate a Power BI app closely by showing Claude the PDF (!) of the dashboards and then made it implement the more sensible filtering of Qlick. 0101 10
Reposted by Karsten Konrad 🇪🇺William B. Fuckley @opinionhaver.bsky.social · 13/10/2026one of the funnier little discursive artifacts is that some people are so pot committed to AI never being clearly useful or helpful that they've twisted themselves into pretending that writing SQL or dealing with health insurance are these wonderful experiences no one should ever rob themselves of 2639433 394
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 15/10/2026We call it training for a reason, not programming. Very different things. 010 1
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 15/10/2026And douse the server racks with holy water afterwards. 030 3
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 15/10/2026My colleagues are mostly business consultants, not developers. We build a sandboxed development environment for them so that they can build small applications safely, and deploy them as web apps. CO, tech stack is all preselected. They use it bc. It’s now a fun experience for them. 020 2
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 14/10/2026What happens on the Klassenfahrt stays on the Klassenfahrt. 1120 12
Reposted by Karsten Konrad 🇪🇺Nafnlaus 🇮🇸🇪🇺🇺🇦 @nafnlaus.bsky.social · 01/10/2026"Stochastic parrot" doesn't just misunderstand LLMs, it also misunderstands literal parrots. 092 9
Reposted by Karsten Konrad 🇪🇺Laurie @laurieontech.dev · 12/10/2026You’ll laugh! You’ll cry! You’ll dance! You’ll cry some more! youtu.be/3TNpOD6bov8?...youtu.beLGTM (Looks Good to Me) - Claude Opus 5.5 music videoYouTube video by Degenerative Pixels 2207 20
Reposted by Karsten Konrad 🇪🇺Gus @gusthema.bsky.social · 12/10/2026Check out this practical developer guide for EmbeddingGemma 2 covering multimodal retrieval across text, code, images, audio, and video, modular encoders (270M to 740M), and Matryoshka truncation (768d down to 128d): developers.googleblog.com/embedding… 3517 51
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 14/10/2026That line in Full Metal Jacket now is more funnier to me. In my native German, a „Stier“ is a bull, so I misunderstood. Sgt Hartman was really dropping two insulting possibilities to Private Cowboy, not just one. 100 1
Karsten Konrad 🇪🇺 @zabong69.bsky.social · 14/10/2026Setting up a listening device this incompetently already answers a lot of questions about what his bosses might say when they talk about him. 210 2