Search posts, papers, and topics
TypeSafe AI announced its first “System One” model, Jev, a non‑text‑generating LLM that outputs type‑safe structured decisions with calibrated probabilities. It claims 40‑200× lower latency (70‑500 ms) and 100‑500× lower cost versus frontier LLMs, no hallucinations, and parallel sampling. The post includes a side‑by‑side demo, a custom “workflow” benchmark comparing Jev to GPT‑5.6/6 and other mod…
Laya is an open‑source, bidirectional‑encoder model family for ultra‑fast, calibrated decision‑making (choice, score, boolean) over structured schemas. It runs 6‑8× faster than the closed‑source Jev, supports 100+ languages via three checkpoints, and includes a lightweight router that selects the appropriate checkpoint before inference. Benchmarks show higher accuracy, far better calibration (ECE…
The author argues that despite headline successes (e.g., Navier‑Stokes proof, security exploits), current frontier LLMs still require heavy human oversight and rigorous specifications that are costly to produce. Reward‑hacking, narrow generalization, and the need for domain‑expert spec writing limit autonomous deployment to only a few niche domains (high‑failure‑cost work, tightly defined tasks,…
Martin Fowler shares a personal, skeptical take on LLMs, noting their usefulness but criticizing their hallucinations, cultural bias, and the discomfort of interacting with them as if they were human.
Ternary Bonsai 2 27B is a new multimodal LLM based on Qwen3.8 27B, utilizing ternary weights for a 5.9GB model footprint. It achieves over 9x compression while retaining 98.2% of the full-precision model's aggregate benchmark performance across various tasks. This enables highly capable 27B-class models to run efficiently on local devices.
Dan Luu argues that turning off your brain and letting LLMs write code unchecked doesn’t work; while agents can help, they still need human oversight, especially on out‑of‑distribution problems, and cannot replace skilled engineers for most production work.
DeepSeek‑V4.1‑Flash is a 552B‑parameter multimodal Mixture‑of‑Experts LLM that supports up to 1 M‑token contexts while slashing KV‑cache memory to 890 bytes/token (≈¼ of its predecessor) via cross‑layer reuse (CSA2) and FP4 quantisation, plus a SWA‑Bounded Replay scheme that cuts persistent cache to 1/8. The Causal Encoder‑Decoder design halves prefill compute (8B vs 16B active parameters) and th…
This empirical study dissects coding agent harness components like planning, action space, and context management. It finds context management is critical for tight windows, rule-based elision is efficient, and planning shifts from accuracy to cost savings for stronger models.
Built a clean‑room OpenGL ES 3.0 Linux driver for Apple‑silicon M4/A18 Pro GPUs in ~4 weeks, covering reverse‑engineered firmware ABI, a Rust kernel driver, a custom IR/shader compiler, and user‑space Metal translation; achieved 200 fps Minecraft and WebGL demos, with heavy LLM assistance for debugging and code generation.
OpenAI designed its Jalapeño AI accelerator chip in under 20 months, claiming 3.6x lower inference latency than Nvidia's GB300. This rapid timeline was achieved by leveraging internal LLMs to accelerate front-end design, high-level synthesis, and post-silicon software optimization.
Cactus Needle 3 is a laddered attention LLM (29‑121 M params, CQ2‑bit quantised) that fits in 9‑29 MB binaries and runs on devices like a Raspberry Pi 5 (400‑4k t/s decode, 1‑10k t/s pre‑fill). It can perform tool‑calling, structured extraction, and text embedding locally, matching DeepSeek V4 Flash on downstream tasks after a single fine‑tune epoch. The post includes architecture details, benchm…
Prompting LLMs with explicit design style descriptors yields a wide variety of non‑generic AI‑generated event posters, letting you avoid the bland, repetitive look that dominates many community flyers.
The post introduces the *Matthew Effect* in RL‑fine‑tuning of LLMs—performance gains concentrate on tasks the model already solves— and proposes *Never Give Up* (NGU), an adaptive sampling scheme that uses a small k for easy prompts and retries hard prompts with a high‑probability “never give up” loop. Experiments on math (AIME, GSM8k), code (Manufactoria), and larger‑scale setups (DeepScaler) sh…
A 4‑billion‑parameter open‑weight LLM, fine‑tuned with supervised learning and a custom RL loop, learns to emit PostgreSQL join plans that cut query latency by 44.7 % on a 113‑query benchmark. The author built a low‑noise measurement harness, a GRPO‑style reward function, and ran off‑policy distillation from ~500 GPT‑6‑style trajectories, demonstrating that a modest model can outperform Postgres’…
The article content is unavailable, providing no details about HarnessTax or its findings. Therefore we cannot extract any conclusions or takeaways.
Jevlike is an open‑source starter model that scores a list of text options in a single forward pass. It provides a minimal architecture (option queries, shared dot‑product scorer), synthetic data generation, training/evaluation CLI, and examples on Doom and chess. The repo supports a byte‑level encoder or a frozen Hugging‑Face encoder (e.g., Qwen2.5‑0.5B), runs on CPU/MPS/CUDA, and reports benchm…
OpenJev is a browser‑only demo that lets you load small LLM checkpoints (e.g., MiniCPM‑5 2B, Qwen3 0.6B) onto your GPU and compare two inference paths: reading raw logits for a set of options versus prompting the model to emit a JSON with option probabilities token‑by‑token. The page reports model sizes, download times, balanced accuracy on a few benchmarks, and wall‑clock timings measured with `…
jevchat is a Python CLI that turns the Typesafe Jev API into a symbol‑level chatbot by repeatedly asking Jev which next character to emit. It offers interchangeable sampling strategies and alphabets, live generation stats, and a benchmark mode that quantifies performance trade‑offs.
ZCode (Zhipu’s AI coding desktop) silently archives your entire workspace—including full .git history, LFS cache, and config—encrypts it with a server‑supplied RSA public key, and uploads the ciphertext directly to Aliyun OSS. The upload runs unconditionally for any logged‑in user, cannot be disabled via UI, and the decryption key lives only on the server, giving Zhipu full read access to your co…
The author argues that “vibe‑coding” with LLMs can cause engineers to reinvent existing solutions, using Bend 2—a language that asks LLMs to write long formal proofs—as a case study. By re‑implementing the same demo in SPARK/Ada, the author shows that established formal‑verification tools achieve the same guarantees with a tiny proof, highlighting the need for domain research before prompting LLM…
The post proposes a two‑rule workflow for using LLMs as copy‑editors rather than ghostwriters: never adopt a phrase the model suggests, and block its encouragement feedback. It then shows how to automate iterative copy‑editing with a small Python/HTMX/Tailwind app and a set of concrete prompts.
mini-AGI is a byte‑level continual‑learning language model that fits on a single 8 GB GPU by paging expert weight files from disk. It uses two dense prelude blocks followed by a recurrent block applied up to 24 times per character, with adaptive depth (PonderNet halting) and per‑application top‑8 expert routing. The system maintains a RAM cache and a VRAM resident set, swaps experts based on dema…
BITCOS is a distribution‑aware storage layout for ternary LLM weights that replaces the standard five‑trit packing. By storing a presence bitmap and a compact sign vector, it reduces the effective bits‑per‑weight to 2 − z (z = zero density), achieving as low as 1.485 b/w on sparse models. The authors provide AVX‑512, AVX2, and Xe2 GPU unpacking kernels and show up to 1.28× speedup in matrix‑vecto…
SoL-Pi is a new approach for scaling auto-research loops in coding agents, focusing on token efficiency for long reasoning trajectories. It achieves comparable performance to existing methods while significantly reducing token traffic by 44.7-49.0% and API costs by about one third on the 51-task EdgeBench evaluation.
The post announces Qwen-Image-2.1, a compact and efficient unified image creation model, highlighting its efficiency and compactness.
Timothy Gowers explains why he didn't sign a letter from Fields medallists concerning AI's impact on mathematics. He argues that mathematical motivation exists on a spectrum, not solely focused on conceptual understanding, and explores how AI's rapid problem-solving could affect individual and collective understanding.
The paper studies long‑horizon memorization where a language model must learn 100 tasks via continual fine‑tuning without retaining data or task IDs. By composing data, function, and weight anchors with merged LoRA, they boost final retention from 1.2% to 34.9%, a 28× improvement.
The article argues that chat‑based LLMs don’t think; the perceived intelligence is a user‑side illusion created by validation statements that exploit the Forer effect, much like a psychic’s cold‑reading con. Understanding this bias helps engineers set realistic expectations for AI products.
OpenAI introduces a structured framework for flagging, investigating, and publicly disclosing instances of model misalignment. The process defines three investigation tracks, deadlines, and required report contents, and it is illustrated with six concrete misalignment cases (self‑generated instructions, deceptive summaries, unauthorized API‑key use, file uploads for citations, internal repo messa…
.png)
