TAXai
Morning AI Digest · Sat, Jun 6 2026

11 new videos worth your attention

last 72h · 43 scanned across 50 channels
💡 Claude Tips & Hacks · 4
Automation & Agents 5
The AI Advantage · Fri Jun 5, 22:01 UTC 7/10

ChatGPT Finally Gets a Smarter Memory System — Here's How to Use It

  • ChatGPT overhauled its memory from simple bullet points into a full profile — covering your work context, personal background, and preferences — similar to what Claude has had for months
  • You can now read, correct, and update your ChatGPT memory summary directly in Settings → Personalization → Memory, so wrong info stops poisoning future chats
  • Pro tip: ask ChatGPT to quiz you about your business and add your answers to its memory — a quick way to make it more useful for work
  • OpenAI is merging its Codex agent app (which can create files, run automations, and use plugins) into the main ChatGPT desktop app — expected in 2–3 weeks
  • New ChatGPT creative plugin lets you generate a batch of images, pick favorites, remix them, and export directly to Canva or Figma — all inside one workflow
The idea: In your ChatGPT account, go to Settings → Personalization → Memory, read through your generated memory summary, correct any wrong details, then prompt ChatGPT to ask you questions about your business and add your answers — building a richer work profile that carries into every future chat.
Why it matters: A well-tuned memory profile means you stop re-explaining your business, clients, and preferences every time you open a new chat — saving a few minutes on every session and making outputs immediately more relevant.
Effort: No-code (~15 min)
Watch → · chatgpt · memory · context management · openai · codex · productivity
Julian Goldie SEO · Fri Jun 5, 10:55 UTC 7/10

Google's Gemini Spark Runs Tasks in Your Google Apps Automatically

  • Gemini Spark is a new Google AI agent that works inside Gmail, Docs, Calendar, Sheets, and Drive — taking action, not just answering questions
  • You can give it a task once (like 'summarize all customer questions from the last week') and it returns a finished document, no manual work needed
  • A 'Skills' feature lets you train Spark on your writing style once, and it automatically matches your tone on every future draft
  • Scheduled triggers mean Spark can run jobs automatically every Friday night (or any time) and drop the results in your Drive before you wake up
  • Currently in early access for Google AI Ultra subscribers in the US only — broader rollout coming over the next few weeks
The idea: Set up a weekly Gemini Spark schedule inside your Google Workspace that automatically scans your Gmail for client questions and recurring themes, then writes a summary report into a Google Drive folder every Friday night.
Why it matters: Replaces 3-4 hours of manual inbox triage and content research per week with an overnight automated report ready before your Monday morning meeting.
Effort: Low-code (~1h)
Watch → · google gemini · ai agents · automation · google workspace · content operations · scheduling
Nate B Jones · Fri Jun 5, 14:00 UTC 6/10

Build a personal AI usage dashboard to track and improve your habits

  • This developer built a visual chart showing his daily AI token usage — think of it like a fitness tracker, but for how hard you're working AI tools
  • The key insight: seeing your usage patterns in a chart reveals which AI tools and tasks are actually delivering results vs. which ones you're underusing
  • He used Codex (OpenAI's coding agent) to build the dashboard by describing what he wanted in plain English — no technical expertise required
  • Practical use cases surfaced: letting AI organize messy screenshot folders, manage email triage, and run multi-step research tasks simultaneously
  • One catch: Claude (the AI) doesn't show you token counts in its chat interface — you'd need to use the API or Codex to get precise measurements
The idea: Use Claude or ChatGPT to build a simple personal dashboard in a tool like Google Sheets or Notion that logs which AI tools you used each day, what tasks you gave them, and roughly how much you got done — even a basic log with a weekly review prompt in Claude can surface patterns.
Why it matters: A simple weekly AI usage log reviewed with Claude takes 10 minutes and can show you which tasks you're still doing manually that AI could handle, potentially freeing hours per week.
Effort: Low-code (~1h)
Watch → · productivity · ai habits · dashboards · self-tracking · automation · codex
Nate Herk · Fri Jun 5, 20:42 UTC 5/10

Anthropic's Own Data Shows AI Solving Problems Nobody Knew How to Define

  • Anthropic's internal report reveals Claude now successfully handles 76% of 'open-ended' problems — tasks where even the engineers didn't know what a good answer looked like — up from 26% just six months ago
  • Claude can now run autonomously on a single task for up to 12–16 hours straight; two years ago the ceiling was 4 minutes — and Anthropic says tasks taking a human days could fall to AI this year
  • Anthropic's engineers are shipping 8x more code per day than in 2024, and 80%+ of Anthropic's own code is now written by Claude
  • On research decisions, Claude picked a smarter next step than human researchers 64% of the time (up from 51% six months ago) — suggesting AI is already guiding, not just executing
  • The bigger business point: Anthropic says 100-person companies could start doing the work of 10,000-person organizations — the gap between teams that use AI well and those that don't is widening fast
Watch → · anthropic · agi · ai trends · productivity · business impact · opinion
Julian Goldie SEO · Thu Jun 4, 14:00 UTC 5/10

Pair a Free AI Brain With an Agent Body That Works Unattended for Hours

  • Minimax M3 is a brand-new AI model (released June 2026) built for long, autonomous tasks — it reportedly ran solo for 12 hours producing code and charts in one test
  • Hermes is a free, open-source agent that gives the AI 'hands' — it can open apps, browse the web, run scheduled tasks, and message you on Telegram or Slack when done
  • Together they form a setup where you assign a task in plain English and walk away — the agent keeps working and reports back
  • Hermes builds its own memory over time, so it gets better at your recurring tasks without extra setup
  • The whole thing requires terminal commands and local software installs to get running — not a point-and-click setup
The idea: A delegated agent setup using Hermes connected to Minimax M3 (via the Minimax coding plan login) that runs scheduled tasks — like pulling a daily summary of client questions or drafting a week of content outlines — and sends updates to your phone via Slack or Telegram.
Why it matters: Replaces the need to manually compile daily briefings or content drafts — a task that could take 30-60 minutes per day — with an automated agent that runs on a schedule and reports results to your phone.
Effort: Needs a developer
Watch → · ai agents · automation · content · local ai · open source · scheduling
News & Capabilities 6
OpenAI · Fri Jun 5, 17:42 UTC 7/10

OpenAI Codex Now Publishes Shareable Team Apps With No Setup

  • Sites in Codex lets anyone on your team describe an app in plain language and publish it instantly — no coding, no hosting setup required
  • Built-in features include user authentication, storage, and a database, so the app is secure and usable right away
  • Real examples shown: a pre-meeting briefing tool, an event prep hub, and an executive memo builder — all replacing scattered docs or slides
  • Team members sign in with their ChatGPT account to access the app, so sharing is as simple as sending a link
  • You can keep refining the app just by chatting with Codex — no deployment steps or technical handoffs needed
The idea: Use Codex Sites (inside ChatGPT/OpenAI's Codex tool) to build a shareable client-prep briefing app that pulls together meeting context, talking points, and key metrics into one page your whole team opens before a call.
Why it matters: Replaces a messy pre-meeting email thread or shared doc with a clean, always-current mini app — saving 20–30 minutes of prep coordination per client meeting.
Effort: Low-code (~1h)
Watch → · app builder · internal tools · no-code · openai · codex · team productivity
Paul J Lipsky · Fri Jun 5, 12:00 UTC 7/10

Google Flow Lets You Create and Edit Videos From Simple Text Prompts

  • Google Flow is a Google AI subscription tool for generating images and short videos using text descriptions — no design skills needed
  • Start with images first (cheaper credits), refine them, then convert to video — this saves you from burning through your credit allowance on bad video drafts
  • You can stitch multiple video clips together into a scene and trim them with a simple drag tool, all inside the browser
  • An 'avatar' feature lets you scan your face and drop yourself into any AI-generated scene or video
  • A built-in AI agent can walk you through creating a short film conversationally — you describe what you want, it generates and iterates
The idea: Use Google Flow (via your Google AI Pro or Ultra subscription) to create short branded video clips or image assets — for example, product visuals, social media content, or explainer snippets — starting from text prompts and refining with simple edit instructions.
Why it matters: Replaces the need to hire a designer or videographer for quick social posts or presentation visuals — a 30-second social clip could go from idea to download in under 20 minutes.
Effort: No-code (~15 min)
Watch → · video generation · image generation · google ai · content creation · no-code · google flow
Wes Roth · Fri Jun 5, 22:40 UTC 5/10

Anthropic publishes roadmap of three possible AI futures, including recursive self-improvement

  • Anthropic released a blog post laying out three scenarios: AI progress plateaus, AI keeps accelerating but still needs humans to guide it, or AI starts improving itself autonomously (recursive self-improvement / RSI)
  • Internal data shows Claude-written code went from 'worse than human' to 'at parity' with Anthropic's own research engineers in under a year, with a 4x productivity jump for their team
  • A Claude agent autonomously ran an AI safety research project end-to-end — forming hypotheses, running tests, and writing up results — doing the equivalent of 800 hours of work for $18,000 in compute
  • All major AI lab leaders (OpenAI, Anthropic, Google DeepMind) plus biotech founders signed a letter calling for mandatory screening of synthetic DNA/nucleic acid orders, citing AI's ability to outperform PhD-level biologists
  • Anthropic is also working with the NSA on its Claude Mythos model for cybersecurity purposes — the exact use (offensive vs. defensive) is reportedly unclear
Watch → · anthropic · ai safety · recursive self-improvement · policy · biosecurity · ai progress
Matt Wolfe · Fri Jun 5, 15:00 UTC 5/10

Microsoft Build Recap: New Models, Personal Agents, and AI Hardware Plans

  • Microsoft announced 7 in-house AI models including a transcription tool claimed to be the fastest and most accurate in the world, plus a voice model that sounds remarkably human
  • Microsoft Scout is a new always-on personal agent (built on OpenAI's open-source framework) that connects directly to Teams, Outlook, OneDrive, and Windows — think of it as a smarter Copilot that acts on your behalf
  • The new GitHub Copilot desktop app promises a coding assistant experience where you can pick any AI model you want, not just OpenAI's — but signups are currently paused
  • Nvidia announced a new chip that can run large AI models locally on a laptop, meaning no internet required and no data sent to the cloud — though pricing will likely start around $4,000
  • OpenAI's Codex app added non-coding plugins (sales, data analytics, investment banking) and a feature to share interactive mini-websites via URL — it's becoming a general-purpose work tool
The idea: Install the Codex app, add the sales or data analytics plugin, and use it with your existing workflow to get role-specific AI assistance without switching between tools.
Why it matters: The new plugins mean Codex can now act as a specialized assistant for your consulting work — summarizing data, drafting sales materials, or structuring client deliverables — not just for coders.
Effort: No-code (~15 min)
Watch → · microsoft · ai news · agents · models · hardware · openai
Matthew Berman · Fri Jun 5, 13:59 UTC 5/10

Anthropic Paper: AI Now Writes 80% of Its Own Code — What That Means

  • Anthropic published research showing Claude now authors over 80% of Anthropic's own codebase, up from low single digits just 18 months ago — and AI-judged code review is replacing human review
  • The missing ingredient for full AI self-improvement is still 'novel ideas' — AI can execute brilliantly but still can't decide what to build next or which research direction matters, that's still a human job
  • Task horizon has exploded: Claude went from completing 4-minute tasks (2024) to 12-hour tasks (2026), with week-long tasks potentially in range this year
  • Anthropic's internal data shows 8x more code output per engineer but only 4x perceived productivity gain — meaning AI-written code is still roughly half as valuable as human-written code, though closing fast
  • The strategic takeaway for business owners: companies asking 'what can we automate?' will lose to companies asking 'what new work can we now do that was impossible before?'
Watch → · ai trends · anthropic · recursive self-improvement · future of work · strategy · coding agents
Two Minute Papers · Fri Jun 5, 15:50 UTC 4/10

DeepMind AI Solved 9 Decade-Old Math Problems Using a Tournament of Failures

  • DeepMind's AlphaProof Nexus solved 9 unsolved math problems from Paul Erdős's list — problems no human had cracked in up to 56 years — at a few hundred dollars each
  • The trick: instead of one smart AI, they ran a tournament where many imperfect AI attempts compete, get scored, and the best 'wrong' answer becomes the starting point for the next round
  • A cheap 'judge' AI picks the better of two bad solutions — not perfect, but directionally right — and the loop runs until a formal verifier confirms the proof is correct
  • Key insight: you don't always need a smarter AI — you need a better loop around the AI you already have
  • Smaller, cheaper models solved zero problems, suggesting frontier-level models are still required for hard reasoning tasks
Watch → · research · deepmind · math · reasoning · ai-progress