Not finding I need anything other than Deepseek Flash lately. It’s so consistently good. And costs like a penny per task. Running via Ollama Cloud with Claude Cowork as the harness.

1 reply

0 recast

4 reactions

Almost a week in and I’m pretty convinced this is the way to go. Current setup is: Fable 5 → Kimi k3 Opus 5 → Nemotron 3 Ultra Sonnet 5 → Deepseek v4 Pro Haiku 4.5 → Deepseek v4 Flash Sonnet 4.6 → Kimi k2.7 Code All accessed via Ollama Pro Cloud.

1 reply

1 recast

4 reactions

Folks, I don’t really care that your software is vibe coded—so long as it ain’t buggy—but please stop making dark blue/black gradient bg with translucent, glassy/glowing edge, rounded buttons. Same goes for black with a bg image, and bitcoin orange accents everywhere. It’s such a tell. Not just that you used AI. But that you have no fucking taste.

0 reply

0 recast

0 reaction

My Hermes agents use glm-5-turbo for almost everything, GLM-5.1 for the extra brains. Cursor and copilot for development I am spending 10x less than the avg vibecoder, yet working non stop

3 replies

1 recast

4 reactions

fable 5.1 still learning how to manage subagents :/

1 reply

1 recast

4 reactions

0 reply

0 recast

4 reactions

AI self-preferencing in algorithmic hiring tl;dr Using AI to craft your resume leads to better shortlisting rates "The bias against human-written resumes is particularly substantial, with self-preference bias ranging from 67% to 82% across major commercial and open-source models. To assess labor market impact, we simulate realistic hiring pipelines across 24 occupations. These simulations show that candidates using the same LLM as the evaluator are 23% to 60% more likely to be shortlisted than equally qualified applicants submitting human-written resumes, with the largest disadvantages observed in business-related fields such as sales and accounting." source: https://arxiv.org/abs/2509.00462 https://x.com/heynavtoor/status/2048088874686300431

0 reply

0 recast

7 reactions

Interesting and unexpected product announcement from OpenSubtitles: an AI-feature-packed media player with everything from instant subtitle matching to auto-translation, and even 4K upscaling. I’d love to see this as an SDK/plugin that could be built into other players, but it’s interesting nonetheless. https://rayplayer.com/en https://rayplayer.com/en

0 reply

0 recast

3 reactions

Anyone played with Osaurus yet? https://github.com/osaurus-ai/osaurus

1 reply

0 recast

3 reactions

cooking something big

0 reply

0 recast

1 reaction

A comprehensive security reference distilled from 150+ sources to help LLMs generate safer code https://github.com/Arcanum-Sec/sec-context

0 reply

0 recast

9 reactions

A tool that removes censorship from open-weight LLMs https://github.com/elder-plinius/OBLITERATUS

0 reply

0 recast

9 reactions

Teaching LLMs to reason like Bayesians Google compressed a classical symbolic Bayesian model into a neural network via supervised fine-tuning. Generalization: Models trained only on synthetic flight data successfully transferred their probabilistic reasoning to entirely different domains like hotel recommendations and real-world web shopping. This suggests the LLMs internalized general Bayesian reasoning principles, not just task-specific patterns. The right training signal (demonstrations of how to reason, not just correct answers) can unlock capabilities that prompting alone can't. Read more: https://research.google/blog/teaching-llms-to-reason-like-bayesians/ P.S. There has been so much exciting foundational research lately that I'm more convinced than ever that there is not only no wall but that progress will accelerate. Three of many examples: 1. SOAR ("Teaching Models to Teach Themselves"), which shows that an AI model can generate useful intermediate problems (stepping stones) for tasks it cannot itself solve. https://arxiv.org/abs/2601.18778 2. QED-Nano, a tiny 4B parameter model, was trained to write Olympiad-level mathematical proofs that compete with models 50x its size. The key technique: instead of reasoning in one long pass, the model reasons in cycles: thinking, summarizing what it's learned, then thinking again conditioned on that summary. This lets it reason effectively across 1.5 million tokens without losing the thread. https://huggingface.co/spaces/lm-provers/qed-nano-blogpost 3. Recursive Language Models (RLMs): Instead of stuffing everything into the model's context window where it degrades, the model treats its input as an external object it can programmatically slice, examine, and recursively call itself on, handling inputs 100x larger than its context window. https://arxiv.org/abs/2512.24601

0 reply

0 recast

9 reactions

The L in "LLM" Stands for Lying https://acko.net/blog/the-l-in-llm-stands-for-lying/

0 reply

0 recast

9 reactions

On of my favourite tricks is to tell Claude to duplicate every change it makes to claude[dot]md files to agents[dot]md, so that if I ever need to make tweaks using a different agent, it adheres to the same set of guidelines. Works decently well.

2 replies

0 recast

0 reaction