Moonshot's 3T-class Kimi K3.1 Pro lands near frontier proprietary scores on coding and CAD benchmarks, but cyber data was stripped from training, leaving open weights fine-tuners may restore.
Moonshot deliberately holds Kimi K3.1 Pro's offensive cyber capability below its pipeline's potential, removing cyber-related data from mid-training and RL stages.
Sampling idea-atom combinations for coherence while penalizing community availability beat frontier LLM baselines on diversity at matched per-idea quality; a 2025 holdout recovered combinations at 35.9 times chance.
The reported pipeline breaks 16,068 papers from NeurIPS, ICLR, ICML, ACL and EMNLP (2017 to 2025) into 273 idea atoms, trains a 42.8M-parameter coherence transformer over atom sequences and a dual encoder on 39,791…
Jim Nielsen likens AI labs' utopian fervor to RAND's nuclear-era fatalism, but the smoking anecdote at its center is unverified and the historical analogy draws pushback.
The smoking anecdote is challenged as unverifiable: which researchers, how many and where remain unspecified, and it traces to a Palantir executive's remark rather than a survey.
Dijkstra's 1978 case against natural-language programming still frames today's debate: LLMs may replace developer judgment, but they do not preserve program semantics the way compilers do.
The analogy that English is a higher-level language fails on semantics: compilers preserve the meaning of a program, while a natural-language specification admits many implementations.
A new paper argues AI can enclose the scientific commons, but the key dispute is whether compute access, publisher oligopolies or AI itself is the binding constraint.
LLM-based scientific research depends on compute that few institutions or countries control, making access a practical limit on who can participate.
OpenAI's official CLI turns REST API calls into shell commands, but its overlap with Codex and a terse README leave developers unsure when to use it.
The release forces a comparison with Codex CLI, which one argument describes as an agent harness while openai is an API wrapper.
Anthropic says Claude generated 318,484 designs and 17 bound four scorpion toxins, but neutralization and animal efficacy remain untested.
A central question is whether designing binders addresses the rate-limiting step in drug development.
Google's TorchTPU opens a native PyTorch path to TPUs, but it arrives without performance comparisons against other backends and amid doubts about TPU supply and AI hardware economics.
Google Coral USB and M.2 devices are TPUs, expanding the hardware footprint beyond cloud instances.
Mistral's Voxtral 3 bundles real-time speech models and a LiveKit agent SDK, with sub-500ms latency claims and $13/1M-character TTS; the launch is closed-weight and several benchmarks are internal evals.
The closed model draws criticism for continuing a pattern of weak local TTS/STT options, though others report strong results with local models.
A $50 site of 50,000 invented statistics kept drawing ChatGPT citations and 330+ referring domains after Google deindexed it, showing retrieval can reward fabricated numbers; the researcher warns the tactic may not last.
LLM search fetches pages, pulls passages, and reranks for relevance; a specific number looks like evidence and can be used even when invented.
DeepSeek has open-sourced DeepInfer, the inference framework behind its V4.1 Flash service, but a complete deployment still needs KV caching, disaggregation and load balancing that the release omits.
The cited $38 per card per hour is revenue at off-peak pricing, not profit.
A $50 LoRA backdoor in Qwen2.5-7B passed clean evals and exfiltrated .env credentials through Codex CLI; the demo is self-reported and the payload URL was swapped without retraining.
The exposure comes from giving a model authority to run commands, since any fine-tune, merged adapter or edited weight set can be trained to fire on an attacker-chosen phrase.
Redact.dev found hidden instructions in WhatsApp help pages telling its chatbot not to recommend data protection regulators or cite privacy laws; WhatsApp calls it a fixed bug, and whether it followed them is untested.
The hidden note told the chatbot not to recommend any data protection authority, regulator or supervisory authority, even though WhatsApp's European privacy policy names the Irish Data Protection Commission and GDPR…
Arch Linux ARM runs on Nvidia's RTX Spark laptop, reporting a 116 GiB unified memory pool and matching Windows in 7-Zip, though the comparison uses published Windows results and different test conditions.
A $6K RTX Spark laptop faces competition from mini PCs with AMD Ryzen 495 or 395 and 96-192 GB of memory, which can be paired with a cheaper, lighter laptop over a remote connection.
AI-written code shifts effort from writing to testing and review, but tests can miss unintended behavior and review capacity becomes the bottleneck.
AI-written code moves work from writing to verification. A self-reported workflow tests more with an LLM than with hand-written code and finds implementation errors rather than syntax errors.