Weekly AI Updates — September 2026 digest (3 of 5 episodes posted: Sept 2, 9, 16)
The short version
- Gunay grouped ~15 updates into four themes: AI moving inside the apps you already use, cheaper building and testing, narrow specialist agents over one do-everything AI, and AI moving into speech, education and physical robots.
- Builders: Cursor can now create, host and merge projects natively without GitHub; Claude Code's /design mode previews several UI variants before writing code (demoed live); Replit's security agent scans for exposed keys, SQL injection and XSS/CSRF, and its free mode auto-escalates to a paid model only when a task needs harder reasoning - 'auto model-switching as UX'.
- Google's education push: Gemini personalized tutoring from your own notes and slides, a free 1-year student plan, NotebookLM-style search that now reaches the open internet with citations, and live transcription that switches languages mid-sentence in noisy rooms. OpenAI added a teen mode (hints, not answers; parental study hours) and transparent-background GPT Image 2 via API.
- Local and physical: Mac mini M6 from $899 makes private on-device AI practical - Gunay's own case is a law firm scanning IDs and filling forms with no cloud; 2,000+ humanoids competed in Beijing at sprinting, boxing and tai chi. Meta AI on Mac now benchmarks your social accounts against named competitors. Claude builds code from a Figma design shared in a Slack 'code channel'.
- Model news: Claude Fable 5.1 - the name is real (Anthropic's Mythos-class tier), not an ASR slip - claimed ~25% cheaper for typical work, ~45% for heavily agentic work, with agentic-coding scores up; Chinese models (Kimi at $15/$20/$40 tiers, MiniMax, GLM from Z.ai, DeepSeek v4, Meituan's LongCat) as cheap alternatives, some free through Cline.
- Sept 9: the agent wave arrived on every layer at once - a frontier model (Fable 5.1 / Mythos 5.1) that self-debugs across a repo, an open assistant (OpenClaw 2.0) that runs on the subscription you already pay for, a multi-agent 'group chat' (Hermes Pantheon), shareable agent templates (Grok), and an agent that drives a professional app (Astra in Blender). Simran's constant refrain: go to the primary source and run one real prompt. Paul asked about Lovable cost - yes, Fable 5.1 makes edits cheaper.
- Sept 16: Sonal's line between the two frontier models - Fable 5.1 does what you specify, step by step; Astra takes a goal, picks its own tools, replans when the itinerary busts the budget, and keeps working while you fetch an API key - with 'allow once' approvals as the guardrail ('keep AI tools in the boundary'). Next week she deploys the Codex build to Vercel with CI/CD.
- Closing advice: don't chase every model - pick the one update that solves a problem you already have and finish one task with it. Prompting tip: ask for an expert's traits, terminology and best practices first, then have the model write LLM-to-LLM instructions from them, instead of 'act as a marketing specialist'.
At a glance, three clicks deep
Skim here first: the closed row is the glance, open is the study card with the key points and timestamps, and the ↓ link drops to that concept's full write-up below.
01Auto model-switching and AI embedded where you already workThe week's tools stopped asking users to choose a model or a surface: Replit's free mode escalates to a str…›
The week's tools stopped asking users to choose a model or a surface: Replit's free mode escalates to a stronger paid model only when the task needs it and drops back afterwards; Cursor hosts and merges code without GitHub; Claude turns a Figma design shared in Slack into code in the same channel; Meta AI lives on the Mac desktop. The pattern is AI as a default inside existing workflows rather than a destination app.
Cursor can create, host and merge 'origin' repos natively - no GitHub connection needed (0:15, l4456164)
Claude Code /design generates substantially different landing-page variants before building (live demo) (0:18, l4456164)
Replit free mode auto-switches between free and paid ('power'/'max') tiers by task difficulty - demoed by Sam Altman and Amjad Masad (0:38-0:39, l4456164)
Replit security agent scans for exposed keys, SQL injection, XSS/CSRF from a Security and Privacy Center (0:22, l4456164)
Share a Figma design in Slack, tag Claude, receive built code in the channel (1:07, l4456164)
Meta AI on Mac benchmarks your social accounts against named competitors with format recommendations (0:43, l4456164)
02Local AI for privacy-sensitive work (Mac mini M6 and the law-firm case)New consumer chips make on-device inference practical for the cases where data cannot leave the building.›
New consumer chips make on-device inference practical for the cases where data cannot leave the building. Gunay's example: an in-house PC and model for a law firm that scans client IDs and auto-fills forms without sending anything to a cloud service - the reason a $899-$2,000 Mac mini matters to privacy-constrained businesses more than to average users.
Mac mini M6 pre-orders from $899 (12-core CPU, 16-core neural engine, 16 GB); ~32 GB builds near $2,000 (1:00-1:04, l4456164)
Law-firm ID scanning and form filling run fully local as the motivating case (1:05, l4456164)
Chinese models (Kimi, MiniMax, GLM, DeepSeek v4, LongCat) positioned as cheap alternatives; some free via Cline (1:31-1:34, l4456164)
Claude Fable 5.1 claimed ~25% cheaper generally, ~45% for agentic work; agentic-coding score cited 42 -> 55 (1:28, l4456164)
03Personalized tutoring and live multilingual transcriptionEducation and speech were the consumer-facing frontier this week: Gemini builds quizzes and lessons from a…›
Education and speech were the consumer-facing frontier this week: Gemini builds quizzes and lessons from a student's own notes and targets knowledge gaps, ChatGPT's teen mode gives hints instead of answers with parental study hours, NotebookLM-style research now cites the open internet, and Gemini's live transcription follows speakers who switch language mid-conversation in noisy rooms.
Gemini personalized learning from uploaded notes/slides/curriculum; free 1-year student plan (0:27, l4456164)
ChatGPT teen mode: auto-detects likely under-18s, step-by-step hints, parental study hours (0:24, l4456164)
Gemini Notebook searches the open internet alongside your files, with citations (0:48, l4456164)
Live transcription handles noisy speech and switches English/Spanish/Hindi/Chinese mid-conversation (0:51, l4456164)
Prompting tip: elicit an expert's traits and terminology first, then write LLM-to-LLM instructions from them (1:42, l4456164)
04The agent wave on every layer (Sept 9): Fable 5.1 + Lovable, OpenClaw 2.0, Hermes Pantheon, Grok templates, Astra in BlenderSimran's four themes reduced to one: agents got more autonomous at the model, assistant, team and applicati…›
Simran's four themes reduced to one: agents got more autonomous at the model, assistant, team and application layers in the same week. Anthropic's Fable 5.1 (with a Mythos 5.1 companion) handles multi-step reasoning over PDFs, contracts and tables and refactors whole repositories with fallback self-debugging; Lovable claims 17% better results at 31% lower token cost on hard edits; OpenClaw 2.0 runs on an existing ChatGPT or Claude subscription or a local model and adds shared team sessions, long-term memory and background automations; Hermes 'Pantheon' 0.21 (Nous Research) assembles researcher/coder/editor/evaluator sub-agents in one chat with a human-approved browser; Grok publishes custom agents as one-click templates; 'GPT-6 Astra' built a textured 3D house in Blender from one prompt.
Fable 5.1 in Lovable: 17% better, 31% cheaper on difficult edits - Paul asked and got a yes for existing apps (0:13-0:19, l4469981)
OpenClaw 2.0: your own subscription or Ollama/LM Studio, multi-user sessions, plugins; cheaper and simpler than Manus (0:25-0:35, l4469981)
Hermes Pantheon 0.21: group-chat sub-agents, cross-device shared memory, human-in-the-loop browser; ~5,500 commits (1:11-1:17, l4469981)
Grok shareable agent templates ('open in Grokbot') (0:44-0:47, l4469981)
GPT-6 Astra in Blender: textured snowy Japanese house in ~1 hour (1:27-1:33, l4469981)
05Multimodal and compute (Sept 9): Muse transcription, Gemini agentic video, Perplexity's privacy gate, Pix.new, Nvidia buys Hugging Face, Uber x WayveThe non-agent half of the week: Meta Muse Voice Transcribe (8 speakers, 25+ languages, mid-sentence code-sw…›
The non-agent half of the week: Meta Muse Voice Transcribe (8 speakers, 25+ languages, mid-sentence code-switching, powering Meta AI dictation); Gemini 3.7 Flash's agentic video processing (39% fewer tokens on an 8-minute I/O video, 88% fewer overall at 7% better accuracy) and a week later Gemini 3.8 Flash plus Flash Cyber for vulnerability scanning; Perplexity routing between cloud and local models with a PII-redacting privacy gate and a Coinbase connector (crypto only); Google Pix.new editable posters with object-level edits and Slides integration (Workspace / AI Pro or Ultra only); Nvidia's ~$12.93B Hugging Face deal with a stated neutrality promise; and Uber's London robotaxis on Wayve's mapless driver, 24 Mustang Mach-Es with safety operators, no surcharge.
Meta Muse: 8-speaker diarization, 25+ languages, free tier + paid (0:35-0:44, l4469981)
Gemini 3.7 Flash video benchmarks; 3.8 Flash + Flash Cyber a week later (0:47-0:55, 1:17-1:20, l4469981)
Perplexity hybrid compute + privacy gate; Coinbase connector (0:55-1:02, l4469981)
Pix.new: click-to-edit posters, French to Japanese swap; paid tiers only (1:04-1:11, l4469981)
Nvidia-Hugging Face ~$12.93B, not closed; Uber x Wayve London (1:20-1:27, l4469981)
Tips: Smithery.ai for MCP servers, yfinance, AlphaSignal newsletter (1:02, 1:26, l4469981)
06Goal-directed agents (Sept 16): GPT-6 Astra vs Claude Fable 5.1, demoed in the Codex desktop appSonal's deep dive.›
Sonal's deep dive. Astra (Plus/Pro only, released ~14 Sept) takes a goal, chooses tools and APIs itself, executes, evaluates against the goal and replans; it works on other parts of a task while blocked on an input (generating files while she pasted an API key) and accepts mid-turn corrections ('the folder is empty'). Demo: Codex desktop (Ctrl+Shift+M to pick the model), one goal-level prompt - a 5-day Singapore trip for four within a budget - Astra chose Python/Flask/HTML, installed dependencies, ran it locally, and recalculated when the first itinerary busted the budget. Versus Fable 5.1: strong reasoning and coding on both, but Fable needs the tool named and works in sub-steps; Astra holds one goal end to end. Guardrail: 'allow once' approvals for anything touching your machine, accounts or a live site.
Astra: goal in, tool choice, evaluate, replan; not on the free tier (0:58-1:03, 1:17, l4469982)
Async work during a blocking input; mid-turn steering (1:29-1:44, l4469982)
Codex desktop app, three panels, GitHub connect; Ctrl+Shift+M model picker (1:22-1:26, l4469982)
Singapore trip demo and the budget replan (1:32-1:52, l4469982)
Astra vs Fable 5.1 (1:54-1:58, l4469982); keep tools 'in the boundary' (1:46-1:47, l4469982)
Next week: push to GitHub and deploy on Vercel with CI/CD (1:48-1:49, l4469982)
07AI on the desktop, by voice and by sketch (Sept 16): Gemini for Windows, GPT Live, Images 2.5Gemini's native Windows app (Mac since April;›
Gemini's native Windows app (Mac since April; Windows 3 Sept) opens with Alt+Space beside whatever you are working in, summarises a Gmail attachment without downloading it, keeps a 50-source notebook and saves to Google Keep - but cannot see another app's errors without a pasted screenshot. ChatGPT's 'GPT Live' voice mode holds a real conversation with interruptions (Settings > Voice > Advanced; free and Plus). Images 2.5 adds sketch-to-image, background swaps, text edits and comment-based erase/add, with 2-4 free attempts - the engine behind the '80s flashback' avatar trend.
Gemini Windows app: Alt+Space, Gmail/Drive summary demo (0:17-0:25, l4469982)
Screenshot limitation; notebook and Keep (0:21, 0:28-0:32, l4469982)
GPT Live voice mode setup (0:33-0:43, l4469982)
Images 2.5 features and free limits; the 80s trend (0:44-0:56, l4469982)
Showcase: Roshan's MapMyRun exporter and FFmpeg editor on Render free tier (2:04-2:15, l4469982)
The concepts in full
Auto model-switching and AI embedded where you already work
The week's tools stopped asking users to choose a model or a surface: Replit's free mode escalates to a stronger paid model only when the task needs it and drops back afterwards; Cursor hosts and merges code without GitHub; Claude turns a Figma design shared in Slack into code in the same channel; Meta AI lives on the Mac desktop. The pattern is AI as a default inside existing workflows rather than a destination app.
Cursor can create, host and merge 'origin' repos natively - no GitHub connection needed (0:15, l4456164)
Claude Code /design generates substantially different landing-page variants before building (live demo) (0:18, l4456164)
Replit free mode auto-switches between free and paid ('power'/'max') tiers by task difficulty - demoed by Sam Altman and Amjad Masad (0:38-0:39, l4456164)
Replit security agent scans for exposed keys, SQL injection, XSS/CSRF from a Security and Privacy Center (0:22, l4456164)
Share a Figma design in Slack, tag Claude, receive built code in the channel (1:07, l4456164)
Meta AI on Mac benchmarks your social accounts against named competitors with format recommendations (0:43, l4456164)
Local AI for privacy-sensitive work (Mac mini M6 and the law-firm case)
New consumer chips make on-device inference practical for the cases where data cannot leave the building. Gunay's example: an in-house PC and model for a law firm that scans client IDs and auto-fills forms without sending anything to a cloud service - the reason a $899-$2,000 Mac mini matters to privacy-constrained businesses more than to average users.
Mac mini M6 pre-orders from $899 (12-core CPU, 16-core neural engine, 16 GB); ~32 GB builds near $2,000 (1:00-1:04, l4456164)
Law-firm ID scanning and form filling run fully local as the motivating case (1:05, l4456164)
Chinese models (Kimi, MiniMax, GLM, DeepSeek v4, LongCat) positioned as cheap alternatives; some free via Cline (1:31-1:34, l4456164)
Claude Fable 5.1 claimed ~25% cheaper generally, ~45% for agentic work; agentic-coding score cited 42 -> 55 (1:28, l4456164)
Personalized tutoring and live multilingual transcription
Education and speech were the consumer-facing frontier this week: Gemini builds quizzes and lessons from a student's own notes and targets knowledge gaps, ChatGPT's teen mode gives hints instead of answers with parental study hours, NotebookLM-style research now cites the open internet, and Gemini's live transcription follows speakers who switch language mid-conversation in noisy rooms.
Gemini personalized learning from uploaded notes/slides/curriculum; free 1-year student plan (0:27, l4456164)
ChatGPT teen mode: auto-detects likely under-18s, step-by-step hints, parental study hours (0:24, l4456164)
Gemini Notebook searches the open internet alongside your files, with citations (0:48, l4456164)
Live transcription handles noisy speech and switches English/Spanish/Hindi/Chinese mid-conversation (0:51, l4456164)
Prompting tip: elicit an expert's traits and terminology first, then write LLM-to-LLM instructions from them (1:42, l4456164)
The agent wave on every layer (Sept 9): Fable 5.1 + Lovable, OpenClaw 2.0, Hermes Pantheon, Grok templates, Astra in Blender
Simran's four themes reduced to one: agents got more autonomous at the model, assistant, team and application layers in the same week. Anthropic's Fable 5.1 (with a Mythos 5.1 companion) handles multi-step reasoning over PDFs, contracts and tables and refactors whole repositories with fallback self-debugging; Lovable claims 17% better results at 31% lower token cost on hard edits; OpenClaw 2.0 runs on an existing ChatGPT or Claude subscription or a local model and adds shared team sessions, long-term memory and background automations; Hermes 'Pantheon' 0.21 (Nous Research) assembles researcher/coder/editor/evaluator sub-agents in one chat with a human-approved browser; Grok publishes custom agents as one-click templates; 'GPT-6 Astra' built a textured 3D house in Blender from one prompt.
Fable 5.1 in Lovable: 17% better, 31% cheaper on difficult edits - Paul asked and got a yes for existing apps (0:13-0:19, l4469981)
OpenClaw 2.0: your own subscription or Ollama/LM Studio, multi-user sessions, plugins; cheaper and simpler than Manus (0:25-0:35, l4469981)
Hermes Pantheon 0.21: group-chat sub-agents, cross-device shared memory, human-in-the-loop browser; ~5,500 commits (1:11-1:17, l4469981)
Grok shareable agent templates ('open in Grokbot') (0:44-0:47, l4469981)
GPT-6 Astra in Blender: textured snowy Japanese house in ~1 hour (1:27-1:33, l4469981)
Multimodal and compute (Sept 9): Muse transcription, Gemini agentic video, Perplexity's privacy gate, Pix.new, Nvidia buys Hugging Face, Uber x Wayve
The non-agent half of the week: Meta Muse Voice Transcribe (8 speakers, 25+ languages, mid-sentence code-switching, powering Meta AI dictation); Gemini 3.7 Flash's agentic video processing (39% fewer tokens on an 8-minute I/O video, 88% fewer overall at 7% better accuracy) and a week later Gemini 3.8 Flash plus Flash Cyber for vulnerability scanning; Perplexity routing between cloud and local models with a PII-redacting privacy gate and a Coinbase connector (crypto only); Google Pix.new editable posters with object-level edits and Slides integration (Workspace / AI Pro or Ultra only); Nvidia's ~$12.93B Hugging Face deal with a stated neutrality promise; and Uber's London robotaxis on Wayve's mapless driver, 24 Mustang Mach-Es with safety operators, no surcharge.
Meta Muse: 8-speaker diarization, 25+ languages, free tier + paid (0:35-0:44, l4469981)
Gemini 3.7 Flash video benchmarks; 3.8 Flash + Flash Cyber a week later (0:47-0:55, 1:17-1:20, l4469981)
Perplexity hybrid compute + privacy gate; Coinbase connector (0:55-1:02, l4469981)
Pix.new: click-to-edit posters, French to Japanese swap; paid tiers only (1:04-1:11, l4469981)
Nvidia-Hugging Face ~$12.93B, not closed; Uber x Wayve London (1:20-1:27, l4469981)
Tips: Smithery.ai for MCP servers, yfinance, AlphaSignal newsletter (1:02, 1:26, l4469981)
Goal-directed agents (Sept 16): GPT-6 Astra vs Claude Fable 5.1, demoed in the Codex desktop app
Sonal's deep dive. Astra (Plus/Pro only, released ~14 Sept) takes a goal, chooses tools and APIs itself, executes, evaluates against the goal and replans; it works on other parts of a task while blocked on an input (generating files while she pasted an API key) and accepts mid-turn corrections ('the folder is empty'). Demo: Codex desktop (Ctrl+Shift+M to pick the model), one goal-level prompt - a 5-day Singapore trip for four within a budget - Astra chose Python/Flask/HTML, installed dependencies, ran it locally, and recalculated when the first itinerary busted the budget. Versus Fable 5.1: strong reasoning and coding on both, but Fable needs the tool named and works in sub-steps; Astra holds one goal end to end. Guardrail: 'allow once' approvals for anything touching your machine, accounts or a live site.
Astra: goal in, tool choice, evaluate, replan; not on the free tier (0:58-1:03, 1:17, l4469982)
Async work during a blocking input; mid-turn steering (1:29-1:44, l4469982)
Codex desktop app, three panels, GitHub connect; Ctrl+Shift+M model picker (1:22-1:26, l4469982)
Singapore trip demo and the budget replan (1:32-1:52, l4469982)
Astra vs Fable 5.1 (1:54-1:58, l4469982); keep tools 'in the boundary' (1:46-1:47, l4469982)
Next week: push to GitHub and deploy on Vercel with CI/CD (1:48-1:49, l4469982)
AI on the desktop, by voice and by sketch (Sept 16): Gemini for Windows, GPT Live, Images 2.5
Gemini's native Windows app (Mac since April; Windows 3 Sept) opens with Alt+Space beside whatever you are working in, summarises a Gmail attachment without downloading it, keeps a 50-source notebook and saves to Google Keep - but cannot see another app's errors without a pasted screenshot. ChatGPT's 'GPT Live' voice mode holds a real conversation with interruptions (Settings > Voice > Advanced; free and Plus). Images 2.5 adds sketch-to-image, background swaps, text edits and comment-based erase/add, with 2-4 free attempts - the engine behind the '80s flashback' avatar trend.
Gemini Windows app: Alt+Space, Gmail/Drive summary demo (0:17-0:25, l4469982)
Screenshot limitation; notebook and Keep (0:21, 0:28-0:32, l4469982)
GPT Live voice mode setup (0:33-0:43, l4469982)
Images 2.5 features and free limits; the 80s trend (0:44-0:56, l4469982)
Showcase: Roshan's MapMyRun exporter and FFmpeg editor on Render free tier (2:04-2:15, l4469982)
Tools referenced
| Tool | Coverage | Moment | Context |
|---|
Session materials
Archived locally on V: — click to open. Companion pages link to the LMS.
Action items
Extraction notes
This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.
Transcript corrections applied
| The transcript says | The trainer actually means |
|---|---|
| Claude Fable 5.1 | Correct as heard - Anthropic's Claude Fable 5.1 (Mythos-class tier); the drafting pass wrongly suspected an ASR slip |
| Bert / Bort | Unresolved name of a platform for building small teams of specialist agents with distinct roles |
| Long cut | LongCat (Meituan's open model) |
| Klein | Cline (VS Code coding extension) |
| Zed AI GLM 5.3 | Z.ai (Zhipu) GLM model family |
| Heretic Hub | As-heard GitHub project claimed to strip model guardrails for security testing; unverified |
| Miro AI Academy | Gunay Aliev's academy - name unverified |
| Sable 5.1 / stable 5.1 / Feble / Febble / Endropic | Claude Fable 5.1 / Anthropic (Sept 9 and 16) |
| Mitros 5.1 / Mitel's 5.1 | Claude Mythos 5.1 |
| News Research | Nous Research (Hermes) |
| Veeve / Wave / Vaze | Wayve (Uber's London robotaxi partner) |
| Grogbot | Grok (xAI) agent templates |
| smithry dot a I / by finance / alpha fold | Smithery.ai / yfinance / AlphaSignal |
| Astro / GPT 6 Astra | OpenAI 'GPT-6 Astra' as presented; the name and attribution are as heard - verify |
| Sundren / Opendra / Arthur / Alpha | Simran (Sept 9 mentor) / the host Arfa |
| Alimony Forge | an Outskill community channel - name garbled |
True on recording day — verify before relying
- Cheap-tier news-roundup digest; every product name, price and benchmark is a snapshot circa 2 September 2026 and likely stale within weeks. Mac mini pricing, Kimi tiers and the Fable 5.1 cost/benchmark figures are as shown in the episode's on-screen material, not independently verified. Sept 2, 9 and 16 are in; Sept 23 and 30 append here as they post. The Sept 9 'GPT-6 Astra' and Sept 16 Astra deep-dive report the presenters' claims - release dates, tiers and the OpenAI attribution were not independently verified.