David Zeuthen

Cool agentic repo discovery: I just ran across ACE paper and open source repo, and I really like its utility. ACE (Agentic Context Engineering) tackles a classic problem we all know from ML: manual, intuition driven tuning versus automated, feedback-driven validation. It essentially brings the idea of hyperparameter tuning loops to prompt and context engineering in live systems. The related blog https://lnkd.in/eG59YaCj highlights a few insights I found particularly relevant: ACE models context as an evolving playbook with three modules: Generator, Reflector, and Curator. Each module handles a different part of the search -> evaluate -> select cycle. It explicitly addresses failure modes like brevity bias and context collapse, which show up often when prompts are manually refined without systematic feedback. The reported gains are substantial: improvements across agentic benchmarks, domain tasks like finance, and sharp reductions in adaptation latency and cost. Can’t wait to try it out, especially for setups where prompt quality and stability drift over time! #agenticApp #LLM #promptEngineering

8

Back in the Linux 0.98 kernel days, when you'd get it on CDs, and you'd build computers by dumpster diving corporate refresh cycles, ... Not all CD rom drives were created equal, because protocols were always just suggestions. I had a super awesome fast SCSI ("scuzzy") drive I had gotten from a Philly bank but it wasn't quite compatible with the "standard" SCSI Linux driver. I found that out because Linux failed to boot from the CD Rom So I jumped on a few BBSs and checked The Internet USENIX groups and figured out what the problem likely was. So I had to use a second machine to slightly modify the device driver by guessing what was the stupidest incompatibility problem and then make a floppy disk to boot from so I could read the CD to get Linux to install. And I hated it and never wanted to use Linux ever again. But these days it's even worse because now lousy engineers are pasting over dumb design decisions with "AI" so now we can both rescue antique hardware for fun, and at the same time molest the sanity of other human beings with even dumber Interfaces and incompatibilities. We like to tease the Japanese for using fax machines but maybe they'll survive the coming AI wars by virtue of being so idiotically simple that they are robust. There's a selection event happening right now and I wonder who is seeing it too. Look, the Google guys used Alta Vista and Lycos to build Google. Don't you think the Google of AI is doing something similar? Maybe you even know them as they keep grinding and building as loudly as possible.

7

Model weights are the new binaries, but treating them as simple, passive data files is a dangerous mistake. When loading legacy serialized models in Pickle format, you aren't reading static math. Pickle acts as a stack-based virtual machine. If an attacker has tampered with the weights and overridden the __reduce__ method, simply unpickling the file can lead to remote code execution. While migrating to Safetensors format is major step forward because it becomes a non-executable raw tensor data and JSON metadata, beware of the pitfall: "oversimplified security labels can create misplaced trust". The security of model loading cannot be guaranteed by the file format alone. Don't let the 'magic' of AI blind you to basic supply chain security.

6

The 2026 ‘DeepSeek moment’ only took a second, as they released another banger paper on New Years Eve. LLMs have tradeoffs between breadth (parameters), and depth (layers) resulting in accuracy (signal) loss during training. DeepSeek’s researchers continue to treat memory as a first-class citizen, introducing hyper-connections which: - prevent signal collapse - persist memory across layers -- not resetting every few steps - compounds computation across layers, instead of redundant processing Together these optimizations improve both processing and memory efficiency, which following Jevons Paradox, will drive even more demand for both. Due to the demands of ever more sophisticated agentic workloads, running 24/7 as our work and life digital companions. Make sure your AI infrastructure is ready for this year. WEKA leads the industry on performance and efficiency. We’d love to be part of your AI journey as well!

20

2 Kommentare

QoS: Congestion engineering for synchronized GPUs. In large-scale ML training, performance is gated by the slowest collective step, not by average bandwidth. That’s why many GPU clusters underperform even when links aren’t saturated. The root cause usually isn’t throughput. It’s late congestion signals and poorly scoped losslessness. QoS for AI / RoCE fabrics: * Lossless must be scoped, not global - Most production RoCE designs dedicate one traffic class for RDMA and scope lossless behavior to that class only. - Making the entire fabric lossless just spreads congestion and increases head-of-line blocking. * PFC is a safety mechanism and not the control loop - Priority Flow Control prevents drops, but it reacts after queues are already full. - Used broadly, it can amplify congestion instead of resolving it. - In real AI fabrics, PFC is treated as a last line of defense, not the primary tuning knob. * ECN is what actually stabilizes training - Early Explicit Congestion Notification lets senders slow down before buffers overflow. - This philosophy traces back many years to Data Center TCP (DCTCP), a paper that showed how early ECN marking and shallow queues dramatically improve tail latency in data-center workloads: DCTCP: https://lnkd.in/gzqenyb8 These same principles have since been reinforced by Google’s datacenter congestion-control research (e.g., TIMELY), which demonstrates how early congestion feedback + low queue depth improves P99 latency at scale: TIMELY (Google Research): https://lnkd.in/gY5tqjHA Together, this line of work directly informs how we design QoS and congestion control for modern RoCE-based AI fabrics. * Scheduling matters more than raw priority - One subtle but critical lesson: - You don’t want bulk RoCE traffic in strict priority. - What actually needs priority is congestion signaling, so feedback arrives fast enough to prevent cluster-wide synchronization stalls, without starving control-plane or storage traffic. The KPIs that actually matter in AI fabrics: When evaluating QoS for GPU clusters, we need to look far beyond link utilization: - P99 / tail step time - ECN mark rate (early vs late) - PFC pause frequency and duration - Which queues or links trigger global stalls At companies like Google, NVIDIA, and Anthropic, the challenge isn’t moving bytes fast but moving them predictably, at scale, under synchronization pressure. Bandwidth gets you demos and determinism gets you scale. QoS, in AI data centers, must be part of the distributed systems design and not a checkbox on a switch. #AIInfrastructure #MLInfrastructure #GPUNetworking #RDMA #RoCE #QoS #ECN #PFC #DistributedSystems #GoogleAI #NVIDIAAI #Anthropic

16