Sep 3 · AI Daily

BestBlogs 09-03 morning report rounds up ten AI items: Google releases Gemini 3.8 Flash and Cyber, boosting long-horizon programming and vulnerability-fixing reasoning capabilities.

Tibo posted on X that the ChatGPT desktop app is now his main browser, noting his productivity has never been higher.

Claude Code has released v2.1.259. It adds a managedMcpServers setting for organizations to provide HTTP/SSE MCP servers to all users, a new --permission-prompts none flag to automatically deny permission requests in headless scenarios, and support for GitLab MR (glab) command recognition and claude plugin validate -- output.

TechCrunch Disrupt 2026 will take place October 13-15 at the Moscone Center in San Francisco. The Builders Stage agenda was announced for the first time, featuring speakers such as Gamma CEO Grant Lee and Google VP of Product Robby Stein.

According to sources familiar with the matter, Palo Alto Networks acquired Console—a two-year-old startup using AI agents to automate IT service desk tasks—for $500 million in cash and stock. Official announcements did not disclose terms. Console had raised $29 million and was valued at $157 million before the sale. Its agent features will be integrated into Palo Alto's Cortex platform, letting security teams use natural language to investigate and resolve alerts. This marks the company's seventh acquisition of 2026.

Boris Cherny demonstrated Fable 5.1 making the Claude tag in Slack more useful, enabling it to pull together executive briefing slides on the fly from metrics spreadsheets and other Slack data, while spotting and flagging a vendor report with inconsistent numbers before proceeding. The Claude tag is now available in Slack for Team and Enterprise plans.

The Databricks official blog published a follow-up to Southern Company's SCOUT storm intelligence case study, continuing a previous article on the company's practices to complete the full story of storm intelligence.

This morning report focuses on capabilities, workflows, and final costs, detailing three pieces: Google released the Gemini 3.8 Flash series including general Flash and Flash Cyber, which achieved an internal vulnerability discovery success rate of over 70%.

Two co-founders of the ATV Big Air Tour used ChatGPT Work to run 26 tours across the US, cutting merchandise inventory and restocking time from 2-3 days down to 2-3 hours. Daily automated briefings reduced event information verification from about 8 hours a week to 1 hour, automatically finding errors and drafting correction emails. After AEO automated site structure audits, OpenAI search and user crawler visits grew from 183 to 2,421 over 30 consecutive days, an increase of 1,223% month-over-month.

@thehypedotnews used Three.js to have Fable 5.1, Fable 5, and Opus 5 programmatically reconstruct three Lord of the Rings scenes (Bag End, Barad-dûr, Rivendell) from a single reference image each, all run via OpenRouter.

ProximalHQ released the FrontierSWE v2 ultra-long-horizon coding benchmark, showing Claude Fable 5.1 in a commanding lead.

The author suggests humans first learn individual words and then understand complex meanings through combination and experience—for example, understanding a meteor shower from streaming, stars, and rain. AI, on the other hand, first builds correspondences between tokens and IDs in a massive corpus, then learns relationships between IDs through neural networks to approximate language meaning. The author argues that this pallidness in AI language is predetermined from the start.

User @marcgayle asked Charlie Holtz's team when Muse subscription support will be added to Conductor. Charlie Holtz reposted the question and replied, "is it time to do this?!", without giving specific times or plans.

Broadcom's latest quarterly earnings report showed AI semiconductor revenue reached $16.7 billion, up 221% year-over-year, accounting for about 56% of total revenue of $29.6 billion. The company expects AI semiconductor revenue to reach $115 billion in 2027 and $230 billion in 2028, doubling for two consecutive years. Q4 guidance projects AI revenue at $21.7 billion (up 236% year-over-year) and total revenue guidance at $34.8 billion.

TechCrunch Disrupt 2026 introduced a new Real World AI stage, focusing on the intersection of the digital and physical worlds and the various ways they continue to merge. The headline mentioned the stage will feature elements like Nvidia, robotics, and extinct animals, while the body text only introduced the thematic direction of the stage.

TestingCatalog reported that the GPT-6-Astra model slug has appeared on the OpenAI API, speculating that if it launches tomorrow it will be a major week, and guessing that routing functionality may roll out first.

Meta released the Muse Spark 1.3 AI model on September 2, 2026, designed to extend agent workflows and enhance coding capabilities. Meta Chief AI Officer Alexandr Wang stated its coding abilities surpass GPT-5.6 Sol and match Claude Fable 5.1.

Spanish AI company Multiverse Computing launched Quasar 438B, a reasoning model with 438 billion parameters, scoring 43 on the Artificial Analysis Intelligence Index v4.1.1—the highest among European models included in the comparison.

A paper comparing scripted experts and ACT-trained imitation policies in the ParcelStow contact-rich task across different speed multipliers found that both achieved 100% success at nominal speed. However, at maximum demonstration speed, the expert success rate was 84% while ACT reached only 53%, with two ACT policies initialized with different parameters dropping by 34 and 48 percentage points respectively, compared to only 16 percentage points for the expert.

The study found that forward KL distillation during the pre-training stage improves both reasoning and factual memory, but during the mid-training stage it slows down the acquisition of factual memory while reasoning continues to improve. The authors propose Switch Distillation, using teacher prediction entropy as a routing signal to distill only on tokens where the teacher is confident, falling back to cross-entropy for the rest.