LINK 🌞 @Lin_______k

@witcheer This model is fast, but when used in real projects with Hermes, it creates many bugs and enters infinite loops. I've reverted back to the Qwen 3.6 MTP standard.

@Taniyatweets_ Regex

@ItakGol Europe has AI ACT : eur-lex.europa.eu/eli/reg/2024/1…

@naasking @KyleHessling1 After 150000 token the model act very weird he entered an infinite loop in hermes.

@weswinder

AMAZING! Claude Fable 5 one-shotted a beautiful geometric islamic patterns visualizer app you could play with. Every seed generates a new pattern. Download any pattern for free.

DFlash can be faster than MTP. But there is no universal winner. It depends on the model, the hardware, the task, the workload, and many other details. I benchmarked MTP vs DFlash with vLLM and llama.cpp across math, coding, and chat workloads. For Qwen3.6 27B, DFlash delivered up to a 4x speedup. MTP was not far behind, once tuned, but DFlash reached the highest peak performance. For Qwen3.6 35B A3B, MTP often performs better. All the results here: kaitchup.substack.com/p/dflash-vs-mt…

keyboard_arrow_left Previous keyboard_arrow_right Next

😂

Justine Moore @venturetwins

2 weeks ago

Me using Claude Opus 4.8 to rename a file

2K 9K 76K 44.3M 10K

Be Chinese AI models: - Cheapest subscriptions - Open-source everything - Let users decide what to use - Strong coding & reasoning - Runs locally on Ollama, vLLM & llama.cpp Meanwhile others are busy locking features behind paywalls.

DeepSeek @deepseek_ai

3 weeks ago

We are making our discount permanent! 🎉 Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life! 🚀

1K 3K 24K 6.7M 6K

🚨💪 “NO HR TEAM, NO PROBLEMS” - BOLT CEO FIRES ENTIRE HR TEAM! “We had an HR team, and that HR team was creating problems that didn’t exist. Those problems disappeared when I let them go.” — Ryan Breslow hails that firing the entire HR team eliminated unnecessary problems created by HR itself! It also helped get rid of “complain culture” at work and go back into execution mode. So apparently removing the people that play thought police at work helps workers work.

@stevibe @UnslothAI 👌

Une prière s’est tenue pour les Palestiniens tués par des colons et des soldats israéliens en Cisjordanie. Le prêtre orthodoxe aux côtés de ses frères musulmans en signe d’unité a souligné l’unité nationale des Palestiniens, musulmans et chrétiens..

Sci-Hub is an evil website that pirated 85M+ research papers and made them freely available And now they've added AI to their database to make Sci-Bot. It answers your questions using latest, full-text articles. But DO NOT use it. We should all try to make billion-dollar academic publishers richer. I'm putting the link below so you know how to avoid it.

Okay this one is insane. A new 18B frankenstein model was just released on @huggingface — Beats the new Qwen3.6-35B-A3B on a 44-test suite despite requiring 12GB VRAM instead of 24GB 🤯 Runs on a SINGLE RTX 3060 (!) 🧠 Opus 4.6 & GLM-5.1 reasoning in one model ⚡️ 66+ tok/s stable on mid-range GPUs 🧪 Experimental, no additional training 🛠️ Perfect tool calling & agentic reasoning 📷 Fits on low hardware, any 12gb card 📚 GGUF size is 9.8GB (Q4_K_M) Another gift from Jackrong, adding both qwopus and glm-distilled qwen together was not on my bingo card. Truly seems like the sweet spot between 9B and 27B models right now. The ultimate model for 12-16GB VRAM owners? huggingface.co/Jackrong/Qwopu…

Anthropic pays engineers $750,000+ a year to understand how LLMs work. Stanford just put a 2 hour lecture that covers 80% of it for FREE. Bookmark this. Give it 2 hours today. It might be the highest ROI thing you do this month:

Warmup to Statistical Mechanics What Exactly is a Hamiltonian A System? In ordinary Mechanics, you might begin with position and velocity. Hamiltonian Mechanics rewrites the same motion in a different language. Instead of position and velocity, it uses position and momentum. We write the position variables as q and the momentum variables as p. Then the full state of the system at one instant is (q, p) That pair is one point in phase space. Why do we do this? Because in these variables, the equations of motion take a remarkably clean form. Everything is generated by one single function, the Hamiltonian H(q, p) and in the simplest cases this Hamiltonian is just the total energy written in terms of position and momentum. So if you know H, you know the dynamics. You might wonder, but how can one function generate motion? The rule is dqᵢ/dt = ∂H/∂pᵢ dpᵢ/dt = −∂H/∂qᵢ These are Hamilton’s equations. Now read them slowly 😄 The rate of change of position comes from differentiating H with respect to momentum. The rate of change of momentum comes from differentiating H with respect to position, with a minus sign. This constitutes the whole engine. A simple example makes this less abstract: Take one particle of mass m moving in a potential V(q). Then the Hamiltonian is H(q, p) = p²/(2m) + V(q) The first term is kinetic energy. The second term is potential energy. Now apply Hamilton’s equations. First, dq/dt = ∂H/∂p = p/m So momentum tells you how position changes. Second, dp/dt = −∂H/∂q = −dV/dq Thus, momentum changes because of force. If you now combine these two equations, you recover ordinary Newtonian mechanics. Since p = m dq/dt, we get m d²q/dt² = −dV/dq So, Hamiltonian mechanics is not a different theory. It is the same mechanics, written in a form that exposes its geometric structure much more clearly. The animation The full 3D surface is the Hamiltonian itself, the energy landscape H(q, p). The floor underneath is phase space, marked by energy contours and the local flow field. The bright moving point is one actual state (q(t), p(t)) evolving under Hamilton’s equations. Its trail shows that the motion is not arbitrary. It is guided everywhere by the geometry of the same single function H. The render is doing more than illustrating a particle moving, it is showing how one function organizes the whole phase-space motion. The math breakdown: Start with one degree of freedom. The state is described by position q and momentum p. So the system lives in a two-dimensional phase space with coordinates (q, p) Now choose a Hamiltonian H(q, p) Think of H as the energy function. In many standard systems, H(q, p) = kinetic energy + potential energy For a particle of mass m in a potential V(q), this becomes H(q, p) = p²/(2m) + V(q) Hamilton’s equations say dq/dt = ∂H/∂p dp/dt = −∂H/∂q Now substitute this specific H. First compute the p derivative: ∂H/∂p = ∂/∂p (p²/(2m) + V(q)) = p/m So dq/dt = p/m Now compute the q derivative: ∂H/∂q = ∂/∂q (p²/(2m) + V(q)) = dV/dq So dp/dt = −dV/dq These two first-order equations completely determine the motion. Now, connect this back to Newton’s law. From dq/dt = p/m we get p = m dq/dt Differentiate both sides with respect to time: dp/dt = m d²q/dt² But Hamilton’s second equation gives dp/dt = −dV/dq So , together they imply m d²q/dt² = −dV/dq This is exactly Newton’s second law for motion in the potential V(q). Thus, Hamilton’s equations do not replace mechanic, they reorganize it. #HamiltonianMechanics #PhaseSpace #ClassicalMechanics #MathematicalPhysics #DifferentialEquations #Mathematics #Physics

And today’s popular music: “blockade” by Trump.

Israel destroyed entire town of Al-Khiyam in southern Lebanon within a month — Al Jadeed News

low quant weights make the embedding model lose all discriminative power. I plotted the cosine correlation matrix of jina-v5, and one can see that low quant makes the model really blind. The off-diagonal similarities are pretty high on Q1/2/3, meaning everything looks similar in the semantic space. Q4 is a sweet spot where model quality becomes acceptable.

Apple has removed Lebanese village names in Southern Lebanon. As Israel invades, they are already setting the state to justify occupation. I’ve never seen something like this.