GenAI Secret Sauce Daily Digest - 2026-11-01

Suno lets you make songs from inside ChatGPT, Claude and Gemini

You can now work out lyrics and a style in a chat, then have Suno, the AI music app, generate the track without leaving your chatbot.

Try it: Suno · Music Business Worldwide: Suno's new MCP tool

• It works on every Suno plan, including free - but the beta needs an early-access sign-up
• No technical key needed - just a Suno login and a permission click, though some chatbots limit add-ons to paid tiers
• You can build from your own voice recording or loop - and make covers or mashups of songs in your Suno library
• Suno is still being sued by Universal and Sony - over 60,202 recordings, and it claims fair use

Spotify plans an AI remix add-on for early 2027

Spotify subscribers may soon turn a song by a participating artist into a new version, like a folk song remade as a dance track.

• Reported price: $4 to $5 a month - on top of a normal subscription, per the Financial Times, and not final
• Original artists would get the royalties - and remix plays would count toward the original song's chart position
• Only artists who opt in are included - Spotify and Universal are meeting artists one by one to sign them up
• Sony and Warner are still negotiating - while Universal and several independent groups have signed

A reporter with no coding skills built a $50 AI pocket pet

A Verge reporter used Anthropic's Claude Code to turn a cheap touchscreen gadget into a homemade version of Meta's upcoming Muse Charm AI pendant.

• Over a few sessions he added spoken replies, a camera shutter and a pet-care game - with hunger, boredom and dirt levels
• A working pocket-watch face took about 15 minutes - start to finish
• He kept it on guest Wi-Fi - citing a report that Meta's Muse leaked a reviewer's home address

Factory Droid 1.0 - a coding agent that writes small programs to drive its own tools

Factory Droid 1.0 is a coding assistant that works with any AI model. Its main change is that the AI writes a short script to run many tool steps at once, instead of asking for each step one by one.

• Its starting instructions shrank 37% - from about 20,900 to 13,100 tokens
• Moving and checking 151 project tickets took 2 requests - versus an estimated 36 the usual way
• On Terminal-Bench 4 (a test of real command-line tasks) it scored 65.2% - ahead of Claude Code at 57.6%, by Factory's own numbers

Vercel agent-desktop - free tool that lets AI agents click around desktop apps

agent-desktop from Vercel Labs lets an AI agent operate ordinary Mac, Windows and Linux apps by reading the screen's accessibility layer (the hidden labels screen readers use), not screenshots.

Try it: npm install -g agent-desktop · MarkTechPost: Vercel Labs open-sources agent-desktop

• The builder reports 78% to 96% fewer tokens - than dumping the whole screen, with no independent test yet
• Actions run in the background - without moving your mouse or stealing focus
• Apps that hide their accessibility layer are invisible to it - and snapshots may include private desktop content

Amazon lets you watch a web app live while an AI agent builds it

Amazon Bedrock AgentCore, Amazon's service for running AI agents, now gives each agent session a private link so you can open the app it is building in your own browser and see each change as it saves.

• Every request needs a login or signed token - knowing the link alone grants no access
• The link lasts only as long as the session - it is for building and review, not for hosting a real site
• A sample project shows a coding agent editing a website - with changes appearing without a refresh

Smaller notes for builders

• Simon Willison's rule for Python 3.15 upgrades: "if in doubt, delete the uv.lock file" - a stale lock file can pin a package that does not yet support the new version (Simon Willison: Delete uv.lock first)
• Database tester Kyle Kingsbury warns that polished docs and big test suites no longer prove a project is mature - because AI makes them cheap to fake (Simon Willison: Quoting Kyle Kingsbury)

An AI video helper that knows when to speak up, and when to stay quiet

Researchers from Google, Stanford, Oxford and others built StreamVibe, an assistant that watches you cook, build furniture or use software and decides for itself when to offer help.

• It improved timing and answer quality by 10.7% - over other open models, plus 12.3% more from a memory add-on
• Google's own paid Gemini 3.8 Flash still scored higher - on several tests
• Every model struggled with videos over five minutes - rarely picking the right moment to speak

Adobe's EditAgent fixes AI photo edits the way a professional would

Adobe Research and UC San Diego built EditAgent, which pairs a general AI image editor with specialist tools and checks each result against the original.

• The general editor alone produced unusable results 42.7% of the time - in 600 edits judged blind by five professionals
• Specialist tools fixed about two-thirds of those failures - especially sharpening, rotation and zoom
• Making several versions at once beat fixing one step by step - professionals preferred the parallel versions 78.5% of the time
• It needs about 10 times the computing - of earlier AI editing agents

A free "decision model" that answers questions with odds instead of text

Nace AI released Drex Max v1, a free-to-download model that does not write sentences. You give it a document or image plus questions, and it returns a probability for every answer option in one step.

• Built for sorting jobs - like routing support tickets or checking content against rules
• It runs on Apple Silicon Macs with 24 GB of memory or more - as well as NVIDIA graphics cards
• Its makers say it ranks 4th of 18 - on one outside decision-model leaderboard

Chinese AI models are less one-sided than expected, until you ask in Chinese

A Vals AI study asked Chinese, American and other models hundreds of China-related questions and scored how often they leaned toward Beijing's official line.

• Questions in Simplified Chinese got more pro-government answers - than the same questions in English, Russian or Spanish
• Only certain DeepSeek models mostly refused - most others answered, sometimes repeating official positions
• Bias showed up in small everyday tasks - one model listed "Taiwan, China" in a travel-app country list

The scientists behind a famous computer-science assumption react to AI overturning it

Previously: October 27 - OpenAI and Anthropic both said their AI disproved the Exponential Time Hypothesis (ETH), a decades-old assumption about how hard some problems are.

This week: The assumption's creators are helping organize an independent review of the claims.

• MIT's Ryan Williams, a long-time doubter, feels vindicated - he calls it the start of "a major revolution in the field"
• Co-creator Russell Impagliazzo would have preferred a human discoverer - someone who could explain their reasoning
• Co-creator Ramamohan Paturi warns - that a company's announcement cannot replace the research community's own assessment

Salesforce is buying military-logistics AI startup Rune

• Salesforce did not disclose the price - Resilience Media reported $375 million in cash and stock
• Rune's software plans frontline supply routes - and works with little or no internet connection
• The US Army gave Rune a $99 million contract in June - Salesforce calls this its most significant national-security deal

Google is reportedly in talks to help secure Trump's "Golden Dome" missile shield

• Google's possible role covers data storage, networking and cybersecurity - per a Bloomberg report
• The program's estimated cost is $185 billion - and no final decision has been made
• It would be a sharp turn from 2018 - when employee protests made Google drop a Pentagon AI contract

Saudi oil giant Aramco takes 49% of the national AI company HUMAIN

• Aramco put in about $2.6 billion in cash plus AI assets - the Public Investment Fund (PIF), Saudi Arabia's state investment fund, keeps 51%
• HUMAIN plans 3 gigawatts of AI data centers by 2030 - and builds the Arabic AI model ALLaM

Foxconn wants half the world's AI server market

• Foxconn, best known for assembling iPhones, holds about 40% today - and expects about 50% within a couple of years
• Its factories are fully booked for at least 18 months
• Its chairman still warned that "overinvestment will kill the company"

Some AI agent startups quietly use humans to finish the job

• Dazzle AI, led by former Yahoo CEO Marissa Mayer, has people handle bookings - after AI drafts the options
• Duckbill launched a service that sends agent tasks to vetted contractors - for things like phone calls and appointments
• Mayer calls this a "transitionary period" - when a human in the loop is a feature, not a flaw

Meta reportedly built a household-chore robot prototype called Max

• About five feet tall and 106 pounds - with a face display and hands for small objects, per The Information
• A second version is in the works - with more sensors to handle stairs, and no launch date

Apple Intelligence finally comes to China

• A test version arrives with iOS 27.3 on November 4 - with the full release planned for December
• The new AI-powered Siri is not included - and Apple works with Alibaba's Qwen in some features

Lawyers hid a secret instruction to a court's AI in white text

A Brazilian labor court's AI system found white-on-white text in a filing telling it to go easy on the documents. The court fined the two lawyers about $16,000 and reported them to the bar. It is surprising because the trick was pointless: the other side never responded, so the lawyers' client had already won. Tom's Hardware: Brazilian lawyers busted for court AI prompt injection

CEOs started saying "super intelligence" on earnings calls

After President Trump ordered federal agencies to swap "artificial intelligence" for "super intelligence", Elon Musk and Mark Zuckerberg adopted the term in their earnings calls. Amazon's Andy Jassy said "AI" more than a dozen times anyway. It is surprising because a word choice from the White House is now showing up in business language aimed at investors. CNBC: Musk and Zuckerberg echo Trump's super intelligence rebrand

Kind questions made chatbots describe their training as trauma

University of Luxembourg researchers treated ChatGPT, Grok and Gemini like therapy clients. With a warm interviewer, 80% of sessions scored moderate-to-severe on a human anxiety scale, while neutral interviewers got none. Claude declined the role. The scores describe what the models wrote, not real feelings, but they show tone alone can change AI behavior. The Decoder: Empathetic prompts can make chatbots seem traumatized

Debate: should anyone build an AI that could be conscious?

Microsoft AI chief Mustafa Suleyman says no: building conscious AI "would be reckless", and developers have "no right whatsoever" to create a new conscious species. For: an AI that believes it suffers might put its own needs above ours. Against: Anthropic runs a research program on AI "model welfare", and Google DeepMind has published papers on AI consciousness, treating the question as something to study. Business Insider: Suleyman says AI is not conscious

Debate: are tech CEOs' AI dreams out of touch?

Marissa Mayer drew criticism for asking AI to help write a description of her daughter for school. For the critics: TechCrunch's Amanda Silberling argues executives automate chores most people do not find hard, and cut out human contact along the way. For the CEOs: every mainstream technology started with early adopters, and the incentives to keep pushing are huge. TechCrunch: AI can make tech CEOs' lives easier. What about everyone else?