Ilyass

Last August, I was invited to keynote (no honorarium) at a major event organized by Polytechnique Montréal—you know, the campus that no longer allows meat to be sold. I received a note yesterday saying they’ve pivoted and will invite someone else to speak. Here's the deal. Unless you’re left-of-centre or aligned with the Liberals, you risk being sidelined on university campuses in Canada. Full stop. That’s the reality—and Canadians should know it.

@Bent302 @spiritbuun @no_stp_on_snek Any impact on token generation speed for the same context window? Especially curious about qwen3. 5:27B

TurboQuant+ results on my local stack using @spiritbuun's CUDA fork. Credit to @no_stp_on_snek for the research. Baseline vs after, same hardware: - Harmonic 9B (5060 Ti): 65K → 384K context. 11GB → 14.5GB VRAM. Nearly 6x the context for 3.5GB more. - Carnice 9B (Quadro RTX 5000): 65K → 256K context. 11.7GB → 13.2GB VRAM. 4x the context for 1.5GB more. - Qwen3.5-27B (3090): 65K → 128K context. 21GB → 20.1GB VRAM. Double the context, less memory than before. Gemma 4 unsupported by the fork — stays on stock llama.cpp at 32K. Fine, it's a compression subagent. All running q8_0-K / turbo4-V with boundary layer protection. No quality loss, no model changes, no requantization. Zero effort. Just flags. Great job guys. I’ll be testing with my actual use cases over this next week and will update.

finally running local LLM setup: RTX 3090 x Qwen3.5 27B x 82k context x 2 parallel slots x hermes agent for knowledge base/research/admin. one gpu, one linux box, no subscription. slower than cloud, but it works and solves privacy concerns.

@digitalspaz What's the Meshtastic radio for?

日本の皆さん、こんにちは。私は、常識破りな銃器技術の発明家です。 こちらが私の最新の発明品、「OptiPew Aimbot（オプティピュー・エイムボット）」です。Bluetoothオーディオを通じてリアルタイムの射撃補正を行う、ステレオ光学式のパッシブ捕捉・シミュレーション装置です。

@Hesamation @huggingface @danielhanchen Keeps getting lost in infinite thought processes for me. Switched back to vanilla Qwen3.5 27B, been getting much better results since and no more infinite reasoning loops.

this model is an agentic treasure. it has been #1 trending for 3 weeks on @huggingface as mentioned by @danielhanchen. it's Qwen 3.5 27B fine-tuned on Opus 4.6 distilled data and beats Sonnet 4.5 on SWE-bench verified and more. "Runs locally on 16GB in 4-bit or 32GB in 8-bit."

We are curious to hear what others are seeing on the degradation side, especially for thermoplastic composites in service conditions. Full publication: sciencedirect.com/science/articl… Learn more: ilylabs.com/materials%20sc…

Key finding: matrix chemistry matters. PPS, PEI, and PEEK degrade differently, different chemical signatures, different surface damage patterns. CF/PPS shows increased cross-linking and brittleness. CF/PEI and CF/PEEK undergo chain-scission under the same conditions.

🧵 How do aerospace thermoplastic composites age? CF/PPS, CF/PEI, and CF/PEEK are increasingly used in aircraft structures. They are lighter than metal, but UV and moisture strongly affects them. We recently focused on longterm exposure to lower altitudes (think zeppelins).

A strain gauge tells you what happened at one point. DIC shows you the whole field. Failure mechanics rarely happen where you expect them.

@sciencegirl Natural fiber composites appear to have real sustainability appeal (if we forget about the matrix). Bamboo has decent specific properties, but the variability is a huge challenge for structural use. Now making carbon fiber out of this would be a real game-changer.

Researchers in China are investigating whether bamboo could help reduce our reliance on petroleum-based plastics. Bamboo is drawing attention because it grows extremely quickly, requires relatively few agricultural inputs, and absorbs carbon dioxide as it matures. Its fibres are naturally strong, making them suitable for reinforcing plant-based polymers. In laboratory settings, bamboo fibre composites have demonstrated mechanical strength comparable to certain conventional plastics, particularly in applications such as packaging and lightweight components. Some bamboo-based biopolymers also break down more readily than traditional plastics under controlled composting conditions. However, moving from promising lab results to large-scale manufacturing presents challenges. Scientists are still studying long-term durability, resistance to moisture and heat, production costs, and how these materials perform outside controlled environments. While bamboo-based plastics are not yet a complete replacement for conventional plastics, ongoing research suggests they could become part of a broader shift toward more sustainable materials.

@MIT_CSAIL Interesting direction. The question is whether ML-optimized composite architectures actually deliver the predicted mechanical gains outside the training distribution. We use DIC to generate dense local field data that could help close that gap. Curious to explore this.

Designing stiff and tough composites is a fundamental engineering challenge. How about using AI? MIT CSAIL researchers present a computational pipeline that combines physical experiments & ML to discover composites w/the optimal balance between stiffness and toughness: bit.ly/4bGfK4a

Got annoyed by Tailscale requiring 3rd party accounts, so created Nostr VPN. It signals over nostr relays and creates a wireguard / boringtun network. Builds for Macos and Linux. Using it between my Macs, but haven't tested extensively yet.

A lot of our experimental work uses Digital Image Correlation. Most materials we study are heterogeneous. Composites, 3D printed parts, biological tissues. Properties vary locally. DIC makes that variation visible. That is why it works so well for us. doi.org/10.6084/m9.fig…

Been meaning to post more about what we're working on in the lab. Upcoming thread series on our main projects: Additive manufacturing: understanding how bead bonding develops in MEAM Thermoplastic composites for the aerospace Heterogeneous materials characterization methods

In theory, meshtastic sounds like a great idea. In practice, I don't know if it is because although I can see nodes, there's no activity at all in my area, so who knows.

This work also showed that high performance thermoplastic composites prefered in the aerospace industry might not be the best choice for an airship, especially when cost and ageing are also considered. Read the full paper here: doi.org/10.1016/j.comp…

These findings provide crucial insights into material behavior under prolonged environmental exposure, aiding in the development of robust composite solutions for aerospace applications like the LCA60T.

The long-term performance and ageing of high-performance thermoplastic composites are critical for their application in demanding environments, such as the innovative structures of next-generation airships.