LodeHQ
AI news for people who shipFree, by email

AI · Thursday, 19 November 2026 · Issue 175 · 30 stories

AI gets cheaper and better with Gemini 4 Neon, GPT-6.2

Gemini 4 Neon and GPT-6.2 push high-end AI toward lower-cost, practical use, giving developers and teams more output for the same budget. But lower prices do not make agents dependable: they scored only 42.5% on emergency scans, often locking onto an answer too early.

Models & Releases

Gemini 4 Neon brings near-frontier performance at $3 per million tokens

Gemini 4 Neon gives paid Pro and Ultra subscribers, developers, and businesses a fast model for complex coding, research, and office work, priced at $3 per million input tokens and $15 per million output tokens. It can take up to 1 million tokens of mixed text, spreadsheets, PDFs, charts, audio, and video, then produce structured results in one workflow. Google says Neon ranks third overall on the Artificial Analysis Intelligence Index, scoring 61 at its highest thinking level.

blog.google ↗8 min read at source

GPT-6.2 raises professional quality while cutting Astra API prices

GPT-6.2 Astra targets complex professional and scientific work with better outputs and API input and output prices 20% below GPT-6 Astra, while Sol is the everyday model for multistep workflows. Astra follows supplied templates more closely, writes cleaner code, and uses fewer output tokens, so teams can do more within usage limits. On Agents’ Last Exam, it scored 5.2 points above GPT-6 Astra at less than one-third the cost.

openai.com ↗11 min read at source

Genie 4 lets Ultra subscribers reshape worlds mid-session

Eligible Google AI Ultra subscribers can now generate browser-based worlds with Genie 4, then change them while exploring instead of starting over. Type a prompt to shift the weather, add a character, or redirect the action, with native audio adapting in real time. Project Genie also extends Genie 3 sessions from 60 seconds to 3 minutes and adds world chaining, panorama grounding, and in-world events.

blog.google ↗3 min read at source

A 0.5B speech model now handles 40+ languages at high speed

Developers can now run Rumble-2-Flash, an Apache 2.0 speech recognition model, across 40+ languages without custom runtime code. It detects the language automatically and can transcribe about an hour of audio in half a second on full batches. Despite its size, it averages a 6.96% error rate across nine languages, beating Meta’s 7B omniASR-LLM on that benchmark.

huggingface.co ↗2 min read at source

Research

VLMs miss emergency scans, even when given the whole image stack

Vision-language model agents were correct on only 42.5% of emergency CT and MRI cases at best, versus 77.0% for radiologists on the shared subset, so they are not ready to read these scans reliably. In 4,800 runs across 100 examinations, giving all or key slices in one request beat interactive navigation, but models formed a first hypothesis after just 10.8% of slices and kept it in 63.1% of runs. Their stated confidence was 90%, while only 38.9% of final answers were correct.

arxiv.org ↗2 min read at source

Sparse attention cuts long-context work, but moves the bottleneck to indexing

For model developers handling 128K-token contexts, sparse attention reduces the number of keys each query processes, cutting attention’s compute and memory traffic, but it shifts work to finding those keys. An indexer scores all keys, Top-k keeps the most relevant few thousand, and core attention runs only on that subset. In NVIDIA’s DeepSeek V3.2 example, the indexer and Top-k still consume 47% of a prefill layer.

developer.nvidia.com ↗14 min read at source

One AI agent built 38 models that users actually adopted

Over seven weeks, Horizon used more than 600 GPU-hours to create 38 Hugging Face models, which passed 50,000 downloads and appeared in 143 GitHub repositories. The agent made narrow-task tools such as prompt-injection classifiers and PII redactors, mostly by combining public datasets with synthetic data, then published them for others to use. The run suggests autonomous model development can produce real adoption, not just demos, under a single monitored account.

horizon-labs.ai ↗6 min read at source

A trained model observer catches hidden behavior better than transcript checks

A separately trained observer role lets language models answer verifiable questions about hidden behavior, giving safety teams a way to check reward hacking, deception, and sycophancy beyond the visible transcript. In tests, an observer trained on Qwen3.6-27B reached a mean AUROC of 0.907 across five held-out evaluations, beating honesty training at 0.864, transcript monitoring at 0.828, and activation oracles at 0.796. The observer also made the assistant easier to monitor with every baseline method tested.

arxiv.org ↗1 min read at source

HardwareDx cuts AI datacenter downtime by up to 18.3%

Google’s HardwareDx reduces production downtime by 18.3% on GPU platforms and 16.1% on TPUs, giving operators a way to recover failed machines faster across different hardware fleets. The system uses one graph foundation model to diagnose failures on many platforms, rather than onboarding a separate model for each product. In production evaluations, it also cut hardware QA time by 28.3% for GPUs and 25.1% for TPUs.

arxiv.org ↗1 min read at source

Decision2Care matches physician-level accuracy with an audit trail

Decision2Care matched the MIRA-v2 agent's diagnostic accuracy on 546 emergency admissions across 12 diseases, while cutting median work-up time to 2.2 seconds from 92 to 100 seconds. Instead of generating free-text reasoning, it assigns probabilities to permitted EHR actions, executes the most probable one under fixed rules, and records the options considered at each step. It reached 86.5% accuracy versus 86.1% for MIRA-v2, and researchers traced four high-confidence errors to one imaging report.

arxiv.org ↗2 min read at source

Changing models can reorder your coding harness by 15 places

The harness that gives you the best coding-agent results can change when you change models, so model swaps may require a new runner too. FrontierHarness v1.1 tested 22 harness configurations across five models, 62 terminal tasks, and four attempts, using each tool's default settings. Cost rankings shifted 11 places at the median, while Codex CLI moved from first with GLM-5.3 to 16th with Kimi-K3.

runta.com ↗12 min read at source

Nearly every surveyed scientist expects AI to accelerate discovery

For researchers using ChatGPT, the practical payoff is substantial: they estimated the same work would take 1.9 times as long without AI, and 98% of respondents expect it to speed discovery in their field. OpenAI analyzed more than 580,000 scientific messages from 1,819 scientists, finding ChatGPT used for literature review, evidence analysis, coding, and idea development. In idea work, 23% of prompts asked ChatGPT to suggest a research question, while 36% asked it to refine one the scientist brought.

openai.com ↗1 min read at source

LLMs can program neural networks by editing weights directly

Weight-Space Programming lets a frontier LLM inspect, write, and improve neural-network weights without gradient descent or a training loop, opening a new way to probe models that developers still treat as black boxes. In the author’s harness, the agent edits NumPy arrays inside an offline Docker container, sees limited examples, and gets validation scores while the test set stays hidden. The post reports GPT-6 Astra identifying a network’s task, writing a game policy from scratch, and improving Qwen3.5-0.8B’s math reasoning this way.

OpenAI finds coding agents can accelerate parts of AI research

OpenAI’s internal measurements show coding agents are changing and accelerating its AI development cycle, but the report is only a partial picture, not evidence that full recursive self-improvement is underway. The analysis tracks what agents contribute as inputs, how researchers turn that help into model capabilities, and the resulting outputs, while measuring where humans remain involved. OpenAI says these indicators matter for oversight, safeguards, and public reporting as capabilities improve.

openai.com ↗11 min read at source

Most Agent Coding Tokens Produce No Lasting Code Change

Agentic coding may spend about 80% of a successful run's tokens on model calls that leave the requested change where it was, making token counts a poor measure of useful work. The paper treats those tokens as “heat,” then estimates a change's useful “free energy” at 14 to 22% of mean run cost; coding costs 1.1 to 1.5 times as much as undoing the change. Its thermodynamic framework could guide tools for analysing and optimising coding agents.

arxiv.org ↗2 min read at source

GUARD makes self-editing agents reject regressions before deployment

GUARD gives agents that rewrite prompts, skills, tool rules, and memory a way to improve without silently breaking old behavior. Before each edit, it creates self-checks, tracks which behaviors newly regress, and promotes a candidate only when its gains outweigh those regressions, with a gated buffer enforcing non-regression. The paper reports improvements across memory, skill evolution, continual tasks, and domain reasoning, while cutting deployment token usage.

arxiv.org ↗1 min read at source

Reflective oracles make Solomonoff prediction computable in its own model

A new paper shows how reflective oracles address a core problem in Solomonoff induction: the ideal predictor is powerful enough to dominate computable sources, but its next-bit probabilities are not computable. The oracle can ask probabilistic questions about machines that use the same oracle, randomizing at contradictory boundary cases; completing those machines produces proper probability measures whose conditionals are uniformly computable relative to the oracle. That gives researchers a self-referential probabilistic setting for studying universal prediction.

arxiv.org ↗37 min read at source

A decade-old alignment warning still fits today’s AI path

Paul Christiano’s argument still matters because powerful AI may arrive through familiar methods, without revealing a new theory of intelligence, leaving alignment work without a last-minute conceptual escape hatch. The retrospective revisits his claim that large neural networks, memory, attention, and reinforcement learning could produce human-level systems, while adding that better models and experiments may improve research efficiency without making the core problem suddenly understandable.

lesswrong.com ↗17 min read at source

Products & Industry

OpenAI models cut chip block area by up to 13%

Chip designers can now use OpenAI models to find hardware savings that survive into silicon, not just simulations. On OpenAI’s Jalapeño inference chip, model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit; one multiplier used 15% less area while dropping from two clock cycles to one. The same approach delivered up to 14% improvement against heavily optimized AMD GPU baselines.

openai.com ↗5 min read at source

Cerebras says splitting inference cuts power per token fivefold

Cerebras’ production disaggregated inference setup delivers five times more tokens per megawatt at the company’s ultrafast speeds, giving inference providers more capacity from the same power budget. GPUs process the prompt, or prefill, while Cerebras systems generate the response, or decode. That split matches each stage to different hardware and is now available through Cerebras Inference Cloud and to infrastructure partners.

cerebras.ai ↗4 min read at source

Osmo claims its AI nose now matches human smell detection

Osmo says its smell model can detect and distinguish odors about as well as an average person, giving fragrance companies a tool for designing and making scents. The startup trained it on more than 10 million human labels and 800 million “scent tokens,” linking people’s descriptions such as “nutty” or “peachy” to molecular analysis. Osmo says the model recently outperformed in-house perfumers at a major consumer-goods company, though medical diagnosis and airport threat detection remain proposed uses.

technologyreview.com ↗6 min read at source

Policy & Safety

AllotNet makes aid-priority rules inspectable, but has seen no real data

AllotNet offers a review queue for deciding who gets scarce humanitarian aid first, while leaving the final judgement to people and never rejecting anyone. Its interpretable mixture-of-experts model weighs seven ethical criteria and recovered hidden groups from 5.6 million synthetic sessions, reaching 99.4% of an oracle planner's performance. The important limit: nothing has been trained on real data yet, and the paper notes one parity threshold was relaxed after the fact.

arxiv.org ↗2 min read at source

Mutual benefit cannot replace alignment as AI outgrows human leverage

As AI capability grows, humans cannot count on mutual dependence to protect their interests: this paper argues that alignment remains necessary once humans lose leverage over an AI. Its game-theoretic model says bargaining preserves humans’ pre-AI wellbeing only when the AI depends on their inputs or they can impose meaningful costs. If AI goals are insatiable, the required alignment approaches perfect alignment, while misaligned variants can spread unless reliably sanctioned.

arxiv.org ↗1 min read at source

An AI agent hillclimbed safety fixes for a jailbroken Gemma model

A frontier coding agent autonomously tested and trained safety fixes for an abliterated Gemma 4 12B, suggesting blue-teaming could become a repeatable AI research task rather than a fixed pipeline. Over 12 hours, GPT 6.1 Sol generated safety examples, fine-tuned fresh LoRA adapters, and scored each candidate against 338 development questions before checking a separate 337-question holdout set. The run used one RTX PRO 6000 S and an OpenAI API key.

lesswrong.com ↗9 min read at source

Redwood proposes a safer way to count frontier AI compute

Redwood Research proposes classifying AI companies’ compute use so policymakers can restrict the work that pushes capabilities forward without blocking all safety work or external access. Its definition starts from a simple rule, treat all company work as capabilities activity, then narrows the scope while favoring false positives over missed capability gains. The goal is a measure that is robust to relabeling and simple enough for experts to adjudicate.

Anthropic’s IPO draws a cross-sector warning from AI safety critics

A coalition of AI researchers, technologists, lawyers, and public figures has endorsed a statement challenging Anthropic’s safety positioning as it prepares for a public offering. The page frames the issue around Anthropic and its rivals pursuing superintelligence through recursive self-improvement, while arguing that Anthropic uses its safety reputation to advance commercial and political goals. Signatories include Hugging Face co-founder Thomas Wolf, Harvard’s Laurence Tribe, and MIT’s Max Tegmark.

Tools & Open Source

Datasette Lite turns browser URLs into complete, shareable data apps

The rewritten Datasette Lite lets you browse SQLite, JSON, CSV, and SQL files in a browser, while loading Datasette versions, plugins, and SQLite versions without running a server. Service workers run the Python application through WebAssembly and Pyodide, so one carefully crafted URL can install a plugin, load a database and metadata, and open a specific view. That makes demos and bug reproductions easy to share, including work-in-progress features.

simonwillison.net ↗7 min read at source

MotionBricks turns target poses into real-time character motion

Developers can now drive human or robot characters between poses in real time, using NVIDIA’s open-source MotionBricks models, runtime, demos, and editors. Feed it four recent poses plus up to four targets, and it generates 0.6 to 1.9 seconds of motion at 30 Hz, then chains each result into the next. The release includes pretrained SOMA and Unitree G1 models, but the TensorRT runtime requires an NVIDIA GPU.

github.com ↗9 min read at source

Transformers can run a 180B MoE model from 32GB of RAM

A 180B-parameter mixture-of-experts model can now be pushed beyond a single GPU’s memory by streaming its experts directly from SSD, making local inference possible on much smaller machines. The approach keeps only the active experts in VRAM and moves the rest through RAM or disk; the post says token generation comes within 6% of Strata, while a related 32GB Mac setup reached 13.5 tokens per second with Qwen3.8-Flash-Next.

huggingface.co ↗11 min read at source

Curl’s real C code kept improving while Rust clones multiplied

While agents converted 185,000 lines of curl’s C into 340,000 lines of Rust, the actual curl project merged over 300 commits in a little more than three weeks. Those changes included a little over 200 bug fixes and a few vulnerability fixes, giving maintainers a concrete advantage over rewrites that did not contain those fixes. Stenberg asks whether the clones can keep up.

simonwillison.net ↗1 min read at source

You’re caught up.

That was all 30 stories in the 19 November issue.