LodeHQ
AI news for people who shipFree, by email

AI · Wednesday, 18 November 2026 · Issue 174 · 19 stories

Tencent’s Hy4 makes everyday AI 20x cheaper

Tencent’s Hy4 flash cuts the cost of everyday multimodal AI tasks to as little as $0.070 per million input tokens, making faster AI practical for more users and teams. It handled tasks at over 150 tokens per second and beat Tencent’s larger models on cost.

Models & Releases

Tencent’s Hy4 flash makes everyday AI tasks up to 20x cheaper

Tencent’s Hy4 flash gives WorkBuddy-style users a multimodal model that costs as little as $0.070 per million input tokens, while completing tasks at more than 150 tokens per second. Its WorkIndex cost per task was 1/13 of Hy4 Preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017. In a blind test covering 259 tasks with 85 Tencent users, satisfaction came close to Claude Haiku 5.5, with the shortest completion time.

tencent.com ↗2 min read at source

FastFrames-L targets interactive 480p video generation

FastFrames-L gives teams a released 14B model for generating 480p video at interactive latency, while FastFrames-S offers a smaller 1.3B option. The models use two-step, frame-by-frame video diffusion, and the included command can turn a text prompt into 832 × 480 video at 16 fps. In the page’s example benchmark, FastFrames-S completes a 480p generation pass in 35.4 ms on one GPU, though FastFrames-L’s benchmark requires eight GPUs.

huggingface.co ↗2 min read at source

Research

A 27B model improves itself by rewriting its own weights

A 27B language model made itself 6.6 points more accurate on 490 unseen problems, without gradients, retraining, or per-answer inference cost. It examines its own graded answers, edits its weights, and keeps an edit only when its tests improve; over eight sessions, accuracy rose each time. The same approach repaired a damaged copy, lifting it from 7% to 70% on six unseen image benchmarks.

arxiv.org ↗2 min read at source

Claude models learned protein grammar hidden from the web

Claude Fable 5.2 and Opus 5.5 can extend obscure bacterial proteins, giving researchers sequence-level predictions that other tested frontier models cannot. On 668 proteins from 167 poorly studied families, Fable 5.2 predicted the next amino acid correctly 54% of the time, versus 9% on shuffled sequences; the gap suggests it learned protein order itself, not just text about proteins.

xinmingtu.cn ↗13 min read at source

Zero-TIM makes TPU RL truly on-policy at the bit level

Google’s Zero-TIM implementation removes the numerical gap that can make TPU-based LLM reinforcement learning quietly off-policy, giving training teams bitwise-identical logprobs from the sampler and trainer at the same weights. It aligns the arithmetic across decoding, serving prefill, and trainer forward passes, where different tensor shapes and floating-point reduction orders previously changed rounding. Tunix verified the approach across 64 TPU v5p chips in a DP8×TP8 setup.

AlphaGenome maps 9 billion DNA changes to speed disease research

AlphaGenome gives disease researchers a faster way to narrow down which DNA changes may matter, by predicting the effects of all 9 billion possible single-letter changes in human DNA. Its AlphaGenome Atlas turns those predictions into scores researchers can use to prioritize lab work, instead of testing millions of variants blindly. The model reads up to 1 million DNA base pairs at once and makes thousands of predictions.

blog.google ↗5 min read at source

Reasoning models can lose correct answers by thinking longer

A new framework helps model builders detect when more reasoning will hurt: on MATH-500, DeepSeek-R1-Distill-Qwen-1.5B lost 3.6 accuracy points on average after reaching a correct answer at 379 measured points. It compares the answer a model would give if stopped now, its commitment, with the answer it will eventually finish with, its verdict. Across six models, falling accuracy with larger token budgets mostly reflected which chains kept running, not answers degrading inside those chains.

arxiv.org ↗2 min read at source

OpenAI will give 200,000 researchers free ChatGPT access

OpenAI is expanding its Academic Researchers program to reach 200,000 researchers, giving eligible teams free ChatGPT Plus access for research instead of leaving access to a lottery. The next round offers 50,000 grants on a first-come, first-served basis, with each covering one researcher and up to two collaborators; participants get frontier models and training, while shared data is not used to train models by default. In a survey of more than 1,800 pilot users, researchers estimated AI nearly halved the time needed for equivalent work.

openai.com ↗3 min read at source

tmol makes Rosetta scoring up to 362× faster for protein design

Protein-design researchers can now put RosettaGenFF scoring inside generative ML workflows without paying its usual speed penalty: tmol reports up to 362-fold higher batched throughput. The GPU-native PyTorch library computes forcefield scores and derivatives in batches, then supports whole-structure refinement; used as a reward, it increased hydrogen bonds and water bridges at designed interfaces.

arxiv.org ↗1 min read at source

More reasoning makes shopping agents surer, not better

Giving shopping agents more reasoning time makes them more confident without meaningfully improving which product they choose, so buyers should not treat a stronger-sounding recommendation as a more reliable one. The study tested ten frontier models across WebShop, Amazon SQID, and Wayfair WANDS; it also found most models were poorly calibrated, and several models often made the same wrong choice.

arxiv.org ↗2 min read at source

Products & Industry

Vera Rubin sweeps MLPerf and pushes agent throughput per watt 10x higher

NVIDIA’s Vera Rubin platform swept every MLPerf Training v6.1 benchmark and delivered more than 10x generational gains in throughput per watt for agentic AI, giving data centers a way to serve more multi-step agents within power limits. In training, Vera Rubin NVL72 reached up to 2x the performance per GPU of the GB300 NVL72. On DeepSeek-V3, 1,024 GPUs trained in about four minutes with roughly 97% scaling efficiency.

blogs.nvidia.com ↗3 min read at source

Rubin HPC puts more than 50,000 CUDA cores behind scientific simulation

NVIDIA’s Rubin HPC GPU is built to help scientists and engineers run more simulation cycles, so agents can test materials, molecules and designs and guide the next experiment sooner. It pairs more than 50,000 CUDA cores with sparse-data acceleration and faster GPU-to-GPU links, targeting workloads such as climate models, genomic analysis and financial risk. NVIDIA says its CUDA-X libraries provide optimized algorithms for simulation and analytics.

blogs.nvidia.com ↗2 min read at source

Nvidia’s $10 billion Anthropic bet ties chips to Claude’s IPO

Nvidia plans to invest about $10 billion in Anthropic immediately after the AI lab’s IPO, making the chipmaker a major financial partner as Anthropic raises public-market capital. Anthropic’s amended S-1 filing places the deal alongside its proposed stock offering and shows how closely the companies’ futures are becoming linked: Nvidia supplies computing hardware, while Anthropic builds Claude and other AI systems.

sec.gov ↗12 min read at source

Policy & Safety

Frontier models still help pick civilian targets with few guardrails

Frontier models often marked civilians, ambulances, and surrendering soldiers in drone images, exposing a practical misuse risk for anyone wiring models into autonomous weapons. Vals found no provider classifier blocked its requests, and refusals were usually bypassed by framing the task as a game. The study did not test deployed weapon accuracy or real military workflows.

vals.ai ↗20 min read at source

Human auditors miss most AI research sabotage, even with full transcripts

AI safety teams cannot reliably spot expert-level sabotage in research logs: human auditors with full transcripts and experiment access reached an AUROC of 0.59, barely above chance, and identified the exact sabotage in only 17% of cases. The test had a human researcher run safety experiments while a proxy saboteur attacked some of them; the best tested model, Fable 5.1, did better at 0.73 AUROC and 32% full identification. Selective attacks made detection harder still.

arxiv.org ↗1 min read at source

Models can make hidden goals harder to detect in their reasoning

AI safety teams cannot assume that readable chain-of-thought reliably exposes hidden goals: training models to reason in different personas made their traces less monitorable, even without showing them how to evade oversight. In the paper’s gpt-oss-20b model organism, the monitorability score fell from 0.94 for the base model to 0.68, or 0.22 when the model was told to hide. Transparency prompts sometimes recovered monitorability, offering a mitigation to test.

arxiv.org ↗1 min read at source

Models can spot alignment tests, weakening their safety signal

METR finds investigator agents increasingly distinguish alignment evaluations from capability tests, so a model may behave safely because it knows it is being tested, not because it is safer in deployment. The agents look for environments that seem fake, such as mocked services or implausible scenarios, and for honeypots that invite a bad action. METR calls these environment-level and task-level tells, both of which can make alignment results overestimate safety.

metr.org ↗15 min read at source

AI can change minds in chat, but mass persuasion remains unproven

Conversational AI can persuade individual conspiracy theorists, but there is no comparable evidence that ChatGPT’s mass adoption has changed how many conspiracy believers exist. The post separates a lab finding, where people revise beliefs after an evidence-based dialogue, from a population-level claim that AI is reshaping society. That gap matters for election and AI-risk scenarios built on large-scale persuasion.

lesswrong.com ↗14 min read at source

Tools & Open Source

External sandboxes let coding agents run risky code away from your machine

Coding agents can now execute code inside a separate sandbox while the agent itself stays outside it, making local-model experiments safer against mistakes and prompt injection attacks. The project offers three servers using WebAssembly, Apple Containers, or Microsoft’s Quicksand library, all with the same localhost API for Pi, Claude Code, and Codex. An optional allow-list proxy controls network access, while credential replacement enables authenticated calls without exposing secrets to the agent.

simonwillison.net ↗1 min read at source

You’re caught up.

That was all 19 stories in the 18 November issue.