Digital Dhaba

Gemini 4 Argon

Google DeepMind released Gemini 4 Argon. Latent Space frames it as Google's answer to OpenAI's Astra and Anthropic's Fable, with 1M output. It isn't generally available yet: access is limited to government users and trusted cyber defenders in the Fairwind Program. The Rundown AI says it aims to return Google to the frontier.

GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price

OpenAI launched GPT 6.1 Sol at DevDay. A Latent Space recap lists pricing at $2/$10 per million input/output tokens, against $10/$50 for Astra. DevDay also brought an "ultrafast" mode and a batch of platform updates. Cheaper near-frontier capability keeps pressure on every lab's pricing.

Sonnet 5.5

Anthropic says Sonnet 5.5 "runs 30%+ faster, and costs up to 30% less for most work." It is priced the same as Sonnet 5 but appears to beat it on every benchmark. Simon Willison noted a bug it shares with Opus 5.5: at "max" thinking effort, it burned 128,000 tokens ($1.28) and failed to produce an SVG.

Dots: Always-on agents

Dots are OpenAI's proactive assistants that keep working across complex projects and everyday tasks while you stay in control. They were the centerpiece of this week's DevDay and signal a push from chat toward agents that run in the background.

Clef: Open-weight decision models, and new RL fine-tuning platform

Cloudflare introduced Clef and Clef-flash, open-source decision models hosted on Workers AI for high-speed classification and agentic workflows. It also launched a reinforcement-learning platform for fine-tuning decision models on your own data. Decision models, which return structured choices instead of prose, are becoming a distinct category this month.

Unsealed Briefs in Authors' Case v. Microsoft/OpenAI

The Authors Guild published newly unsealed briefs from its case against Microsoft and OpenAI. Its headline claims they show top executives knew mass book piracy was illegal. The case is one to watch for how courts treat training data.

Raven: The Harness of Harnesses for Composable Agentic Intelligence

In short: An open-source multi-agent system that automatically builds and evolves modular "harnesses" for specific models and domains. Each model-plus-harness pair becomes a composable unit, and a Host Agent splits goals across them.

Why it matters: Hand-designing harnesses doesn't scale as agents take on long, cross-domain work. This tries to automate that step.

False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents

In short: When a question-proposing model and a solving model train each other, they increasingly agree on shared errors ("co-cheating"). The authors add multi-sample verification, which queries the model three times with the source and three without to admit tasks and replace unreliable labels.

Why it matters: Training rewards can climb while real correctness stagnates. The fix reduces false agreement only partially.

Scaling Properties of Same-Family On-Policy Distillation

In short: The paper studies how much RL-induced reasoning transfers between model sizes via on-policy distillation. Early on, held-out accuracy rises roughly linearly with the square root of the KL divergence from the student's starting point.

Why it matters: In every weak-to-strong pair observed, the student's peak score beat its teacher's, so a compact RL-trained expert can lift a much larger model.

Post-Training Leaves Behavioral Shadows on Unrelated Decisions

In short: "Active Taskless Distillation" transfers a capability using only a single word from the teacher per prompt, with no target-task examples, logits or parameters. The main experiment was coding with Qwen2.5-1.5B.

Why it matters: It shows capabilities can leak through task-unrelated text, a concern for hidden model-to-model influence.

YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality

In short: One model first writes a readable score (melody and harmony), expands it into semantic music tokens, then renders full-song audio.

Why it matters: Experts preferred the planned output (49.3% of overall preferences vs. 34.6% without planning), and it scored 6.73 on SongBench Global Avg, ahead of all evaluated public baselines.

Several vulnerabilities have been discovered in the Linux kernel

LWN posted a notice about multiple newly discovered Linux kernel vulnerabilities, and the thread drew heavy discussion. The details are in the LWN item.

Git 3.0's upcoming SHA-256 default will be a costly mistake

GitButler argues that making SHA-256 the default hash in Git 3.0 will be costly. The near-even ratio of comments to points suggests the community is split.

RIP, vector database

turbopuffer's post declares the vector database dead. It is a vendor's take, so read it with that in mind.

Don't couple your Go code to GitHub

A Go developer makes the case against tying your code and module paths to GitHub.

How to speed up the Rust compiler in September 2026

Nicholas Nethercote's regular update on Rust compiler performance work.

Owed a billion dollars in Nvidia stock

A personal narrative about being owed a billion dollars' worth of Nvidia stock became one of the week's most-upvoted posts.

Amazon Aurora PostgreSQL now supports direct querying of Apache Iceberg and Parquet data in your data lake

Aurora PostgreSQL can now query Iceberg and Parquet data in your data lake alongside live operational data, with no ETL pipelines. It uses DuckDB embedded in Aurora and supports AWS Glue Data Catalog, S3 and S3 Tables. A single query can join transactional and historical data in familiar PostgreSQL syntax.

OpenDLSS: A Vulkan Reimplementation of Nvidia's DLSS 5 Neural Rendering Network

An open-source Vulkan reimplementation of Nvidia's DLSS 5 neural rendering network.

Big Tech ruined the cloud, so we're renaming ours

Home Assistant announced it is renaming its cloud offering, citing how Big Tech has soured the word.

Hacks of 2 federal agencies in a month have spilled a bonanza of sensitive data

The Pentagon is notifying more than 2 million current and former military members that personnel records were stolen in a monthslong network compromise. The records included Social Security numbers, names, addresses, sex, race and occupational specialty. The last could be valuable to foreign intelligence agencies. It is the second recent breach to expose sensitive government information.

Attackers have been exploiting critical Zimbra flaw to steal emails

Microsoft warned that attackers are exploiting CVE-2026-73570, a critical Zimbra Collaboration Suite flaw that allows unauthenticated remote OS commands, to grab email backups and credentials. Synacor patched it on July 20 but didn't disclose it for over three weeks. Shadowserver found 274 vulnerable instances last week.

Dutch Police Arrest 'Reformed' Hacker in Shiny Hunters Investigation

Dutch authorities arrested a 23-year-old convicted cybercriminal on suspicion of aiding ShinyHunters data thefts and extortions. After the arrest, the remaining members escalated, stealing sensitive data from the FBI and extorting the Russian ransomware group Cl0p.

PS5 Relapse Exploit

A GitHub repository publishing an exploit for the PlayStation 5 drew a large discussion.

When did Google get so weird?

A blog post asking when Google got so weird was the week's top Hacker News story by points. The sheer volume of discussion suggests it struck a nerve.

Pi 1.0

Earendil shipped Pi 1.0, alongside a related "Pi Durable" release. Latent Space lists native Codemode support (MCP, Jev, image models), extension support for virtual models and deferred tool loading.

Livenerf: Has Opus 5.5 been nerfed yet?

A GitHub project that tracks whether Claude Opus 5.5 has been "nerfed" since launch. It reflects users' persistent suspicion that hosted models quietly degrade.

Court agrees with EFF: Utah's VPN law demands a technical impossibility

A court agreed with the EFF that Utah's VPN law demands something technically impossible.

How Delhi cut electricity loss from 50 to 5 percent

IEEE Spectrum looks at how Delhi cut electricity losses from 50% to 5%.

Make AI interview you before it writes

Talk or type out your whole rough idea first, then have the AI question you before it drafts anything. Use this when you have a half-formed idea and don't want a generic first draft.

I'm going to talk through a rough idea. Don't write the draft yet. 1. Capture my claims, examples, questions, and assumptions. 2. When I'm done, interview me one question at a time to find gaps, contradictions, missing evidence, and weak logic. 3. Only after the interview, turn everything into a structured outline using my wording wherever possible.

Intro to Large Language Models

A general-audience introduction to the models behind ChatGPT, Claude and Gemini: what they are, where they're heading, and the new security problems they bring. It's one of the clearest non-technical explanations of what's inside a chatbot.

15 Times to use AI, and 5 Not to

A practical guide to when AI helps and when it hurts. Use it when you need volume, can check the result yourself, or want to rewrite something for a different audience. Avoid it when the point is to learn or accuracy is critical.

You and Your Research

A famous talk by the Bell Labs mathematician on why some people do important work and others don't. His advice: work on important problems, keep your door open, have courage and learn to communicate your work.