Halvar Flake
@halvarflake
Choose disfavour where obedience does not bring honour. I do math. And was once asked by R. Morris Sr. : "For whom?" @[email protected]
I will have to print & frame this tweet :-))
If you watch one talk this month, make it this one. Profoundly insightful on the complexity threat to security by @halvarflake given at NATOs CYCON. Video: err.ee/836236/video-g… And slides: docs.google.com/presentation/d…
I spent some time looking at the Vespa source code to see how it compares with Lucene jpountz.github.io/2025/07/25/mor…
Serving Large Language Models on Huawei CloudMatrix384, Huawei & SiliconFlow, arXiv, Jun 15, 2025 x.com/ogawa_tter/sta… 910C Chip => 910C Node x.com/ogawa_tter/sta… May 19 Modular Orthogonal Arch Unified Bus UB APNet 2021 x.com/ogawa_tter/sta… Apr 8 x.com/ogawa_tter/sta…
=> Unleashing the Power of AI, AI Ready Networking & Storage Solution of Huawei, HPC-AI Swiss Conf, Apr 8, 2025 youtube.com/watch?v=3jVFsO… hpcadvisorycouncil.com/events/2025/sw… Unified Bus x.com/ogawa_tter/sta… APNet 2021 UB-Mesh CloudMatrix x.com/ogawa_tter/sta… 384 x.com/ogawa_tter/sta…
Ranked: Per capita emissions from electricity generation in the world's largest power sectors (2024) 🇺🇸 US: 4.9 tCO2e 🇨🇳 China: 4.0 tCO2e 🌎 World: 1.8 tCO2e 🇪🇺 EU: 1.3 tCO2e 🇮🇳 India: 1.0 tCO2e
This is an important point. There's no point for a forward NN pass to hit an avoidable cache miss, the memory access patterns are all precisely known.
Compilers rely on heuristics to compile Turing complete programs, you have to, the runtime is data dependent. NNs aren't like this. They can be directly optimized to maximize execution speed. And it'll be hardware universal. Just burn compute to make it fast. Bye bye $4T moat.
I see the IMO results, and the confidence about LLMs, and I am optimistic about the future, but ... I also spent time reviewing Claude and Gemini C++ code today, and today's performance would have me not hire them. The uneven performance is so weird.
Please note, we're not able to reproduce the 41.8% ARC-AGI-1 score claimed by the latest Qwen 3 release -- neither on the public eval set nor on the semi-private set. The numbers we're seeing are in line with other recent base models. In general, only rely on scores verified by…
I think corruption should be cleaned up everywhere, but Sergej has a strong point that Schroeder has not been touched.
Germans talking about „Ukrainian corruption“ is the most colonialist thing I have ever seen. Like really, guys, we have Schröder not yet imprisoned. Please just shut up.
Madyar and his team are the most experienced and innovative drone warfare unit in the world right now. They’ve redefined the battlefield with tactics NATO hasn’t even trained for. Instead of offering advice, NATO should be sending instructors to learn from them. Drones are the…
Madyar: The next challenge for Ukraine is replacing infantry with ground robotic systems that will replace all logistics on the front line. 4X
Most efficient way I've found to vibe code with Claude Code is to only let it run free on initial application generation or other large fairly-independent tasks For adding or modifying functionality, it saves time to watch and approve every change over correcting things later
The productivity improvement curve for coding with AI agents flattens out *very* quickly as a project grows and tasks move from one-shotting entire files and functions to edits requiring contextual understanding.
Fascinating to compare the solutions of OpenAI vs Deepmind to the IMO 2025. Both won Gold for answering P1 to P5 correctly. OpenAI (left) vs Gemini (right)
Dr. Richard Battin and Dr. Hal Laning were pivotal during the Cold War. Though their efforts were highly classified at the time, their work in compilers, autocode, optimization, and guidance laws certainly helped win the Cold War, and ultimately the Space Race. Which they were…
Everyone's talking about AI performance on the IMO. Let me highlight 🇨🇦Canadian 11th grader Warren Bei🇨🇦, one of five participants with a *perfect* 42/42. This is his *fifth* (and final) IMO representing Canada, with three golds and two silvers. (➡️ MIT undergrad in the fall)
"Epstein, Epstein, alles muss versteckt sein" - slight variation on a German children rhyme.
101 Chrome Exploitation — Part 0: Preface We are starting a new series on modern browsers' architecture and their exploitation using Chrome as an example. Readers will learn how browser subsystems are implemented, how their security is ensured and how it is violated with…
The difference in spreads, and comfort, and UX, and fees, between IBKR and *any* European bank or broker are just stunning. IBKR is so much better in literally every possible dimension. Banks are adding so many shadow fees everywhere it is crazy. Where is the European IBKR?
In early July, Epoch AI estimated that 33 models had been trained with at least 10^25 FLOP. So, here's your REMINDER that the doomers asked to ban/restrict models above 10^21 FLOP (Conjecture), 10^23 (CAIS, ControlAI), 10^24 (GovAI), then 10^25 (FLI, PauseAI, ICFG). They viewed…