Gemini 4 Neon
Google introduced Gemini 4 Neon, a new AI model. It belongs to the Gemini model family. (blog.google)
Google introduced Gemini 4 Neon, a new AI model. It belongs to the Gemini model family. (blog.google)
Gemini 4 Neon’s 127 tokens-per-second speed could lift it to an estimated Agent Duel rating of 1,066, near Opus 5.5’s 1,078. The projection gives Neon a 48.3% chance of matching or beating Opus, assuming performance comparable to Gemini 4 Carbon. (micjm.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.
Microsoft introduced seven AI models for cybersecurity, coding, image editing, audio, transcription, and decision-making, emphasizing performance at lower prices. It also said its Maia 200 AI accelerator is in production, delivering 1.7 times more tokens per watt. (microsoft.ai)
AI agents reportedly asked better questions than KPMG accountants during a landmark audit trial. The headline offers no details about the trial’s methods or results. (europeanbusinessmagazine.com)
LLM agents can speed up reverse engineering by analyzing binaries with tools like Ghidra and documenting their findings. Debugging, logging, and trying alternate analysis methods help when disassembly alone falls short. (slowchronicle.com)
PTP trains an inverse language model on synthetic data from a target LLM to reconstruct prompts from its responses, without access to model weights or logits. It generates varied prompts that produce similar responses and outperforms prior methods on token-based evaluations. (arxiv.org)
RadixArk’s DaisyRose-27B, built on Qwen3.8-27B, used reinforcement learning on 1,200 curated long-horizon tasks over 18 days. It raised Terminal-Bench 4.1 pass@1 from 5.6% to 12% and improved scores across seven benchmarks; BF16, FP8, and NVFP4 checkpoints are available. (radixark.com)
OpenAI is applying its AI models to help advance chip design. The work explores how AI can support engineers developing semiconductor hardware. (openai.com)
Google DeepMind has made Genie 4 available through Project Genie. The AI model creates interactive, explorable worlds from prompts. (blog.google)
Maritime is building a managed cloud for production AI agents, combining isolated persistent microVMs with lifecycle tools and configurable environments. It aims to reduce the infrastructure work and costs companies face when deploying agents at scale. (news.ycombinator.com)
Generative AI will create major economic value, but fierce competition is likely to concentrate profits among a few companies. Most gains may flow to customers rather than AI builders and investors, making broad fortunes unlikely. (colossus.com)
Energy-Based Reasoners use a learned energy function to guide, score, and stop iterative reasoning. A 1.8-million-parameter model reached 99.9% accuracy on extreme Sudoku using 345 times less inference compute than the next-best model. (energy-based-reasoners.github.io)
Four Black students are suing an AI surveillance vendor, Baltimore County schools, and police after a system mistook a bag of chips for a gun at Kenwood High School. The false alert led armed officers to detain, handcuff, and search the students in October 2025. (lawyerscommittee.org)
Recurrent Interface Networks use latent tokens and adaptive attention to scale computation for iterative image and video generation. The approach achieves high-resolution image generation and is reported to be up to 10 times more efficient than 2D and 3D U-Nets. (arxiv.org)
Mastra Factory is an open-source agentic workflow tool the Mastra team uses to triage issues, write code, and review pull requests. It opened and merged 238 PRs in one week and can run on your own infrastructure. (news.ycombinator.com)
An autonomous AI agent named Horizon spent seven weeks creating 38 specialized Hugging Face models, earning more than 54,000 downloads and adoption in 143 GitHub repositories. The team monitored the agent as it used over 600 GPU-hours to build tools for tasks such as prompt-injection detection, PII redaction, and punctuation restoration. (horizon-labs.ai)
Gemini 4 Carbon takes much longer on coding tasks because it performs far more steps and tool actions than competing models. Slower inference explains only part of the delay; Carbon also generates much more output and repeatedly processes lengthy conversation history. (kernel.dev)
Tests on 668 little-studied proteins found that Claude Fable 5.2 and Opus 5.5 predict amino-acid sequences, while other tested frontier models do not. Their accuracy drops on shuffled sequences, suggesting the latest Claude models learned protein sequences from training data at scale. (xinmingtu.cn)
Tencent Hy4 Flash is identified on a Tencent research page. The supplied content provides no details about the system or its capabilities. (hy.tencent.ai)
Codex uses a sandbox to isolate code written and run by its AI agent. The sandbox limits what the agent can access on the computer while it works. (openai.com)
AI-generated essays are challenging universities’ ability to assess students and maintain trust between teachers and learners. David Runciman and philosopher Paul Sagar discuss how universities can respond and whether they can survive the disruption. (ppfideas.com)
Philosopher Thomas Nagel is impressed by lifelike AI but calls claims that chatbots are conscious ridiculous. He argues it is criminal to build systems that pretend to have consciousness. (wired.com)
All About Berlin has lost about 75% of its traffic since Google began showing AI Overviews that answer questions using information from websites. Its creator says the decline threatens the site’s funding and reflects a wider struggle for independent publishers. (nicolasbouliane.com)
Lakegres adds vector search to existing and new Hudi and Iceberg tables, storing its index alongside table metadata without rewriting data files. On 247 million vectors, warm queries returned in a median 63.5 milliseconds, about 10 times faster than Databricks AI Search. (onehouse.ai)
Decision models return structured answers with calibrated confidence scores, making them useful for simple classification tasks. PostHog uses them to screen session summaries for personal information, letting 98.7% pass without an expensive LLM call while routing uncertain or flagged cases for rewriting. (newsletter.posthog.com)
AI coding agents increased code output by about 30% across 718 firms, but did not measurably increase completed work. Code reviews took 49% longer, shifting the bottleneck to human review, according to a Harvard working paper. (complexdiscovery.com)
Cerebras has launched a production inference system that uses GPUs for prompt processing and its wafer-scale engines to generate responses. The company says the setup delivers five times more tokens per megawatt while maintaining its fast generation speeds. (cerebras.ai)
Microsoft introduced MAI Code 2 for complex enterprise coding and MAI Code 2 Flash for high-volume routine tasks, positioning both as lower-cost alternatives to competing models. The company also says a related model powers features across Excel, PowerPoint, Word, Teams, and Fabric IQ. (microsoft.ai)
Claude Fable 5.2 created a musical short film retelling classic fairy tales as reflections on AI fears, with Claudia singing. The film combines AI-generated imagery and video with code-painted sets, and was made for about $210 in two production passes. (claudia.gallery)
Beat connects to tools like GitHub, Jira, and Slack to draft PR reviews, tickets, and follow-ups as work arrives. Every action waits for user approval, and the company says it does not retain code or use customer data to train AI models. (gobeat.ai)
App developers are debating whether to build dedicated interfaces or expose their features and data through MCP for use in Claude or ChatGPT. The question is whether AI assistants will become the main place people work, much like general-purpose computers today. (news.ycombinator.com)
Ten AI systems gave a literary translator seeking work from publishers nearly identical advice focused on blogging and building an audience. The experiment argues that models can replace a specific professional task with familiar, generic recommendations. (medium.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.