Hey gang,

I'm having a splendid time en français (apart from the internet).

note to self: MAKE SURE YOU TRAVEL WITH STARLINK

this is also part of the reason why I've postponed "The AI Course" release.

Doors were meant to open Monday with kickoff on July 6th, but I've moved it back a week. Doors now open Monday the 29th, with kickoff around July 13th.

I refuse to put anything into the world that isn't exceptional, or as close to it as I can get.

A few pieces still needed their final touches: the onboarding flow when you join the program, locking in session times, and finishing the last bits of the curriculum, so it's the most valuable (and digestible) it can be whether you're a total beginner or already running these agents every day.

So pumped to release this in a week's time and get you MFs AI power users!

Here's the waitlist link in case you missed it: aiwithremy.com/waitlist (only 60 spots, and 890 of you on the list now), and waitlisters get the early-bird discount.

📌 TL;DR

• Sakana Fugu → Japan's "new model" is actually a router that auto-picks the best AI for each task. clever idea, but it's slow right now and didn't beat just using one good model. (no Fable 5 in its pool, and it doesn't care.)

• Claude Tag → Anthropic dropped Claude into Slack as a teammate you @mention. it remembers context, takes real actions, and in ambient mode acts on its own. Team/Enterprise only, but the pattern's worth stealing.

• Perplexity Brain → a real memory for your AI agent. it remembers the WORK, not just your preferences, and rewrites its own playbook overnight.

• Builder's notes → the /teach skill that turns any folder into a tutor, how I finally nailed local vs cloud scheduled tasks, and Helm, a little VS Code extension I vibe-coded.

Sakana Fugu, the AI that picks your AI for you

Japan's Sakana AI dropped Fugu this week, and the internet instantly branded it a "Fable 5 killer."

It's not. and don't let the hype fool you, Fugu isn't a new frontier model at all.

It's a router (basically a smart switchboard). you hand it a task, and instead of you guessing "hmm which AI should I use for this?", it auto-picks the best model for the job, Opus, GPT-5.5 or Gemini, then fuses their answers back into one.

Under the hood it's a small trained "conductor" that calls other labs' top models behind a single API. nothing to download, it's hosted only.

Clever idea. not a brand new one, mind you, Perplexity's agent ("Computer", the thing that goes off and does tasks for you) already routes like this. but it's a cool concept.

In real testing though, the cracks show…

Jack Roberts (big name in the AI space) put Fugu up against Opus 4.8 and GLM 5.2. about 3.5x slower than Opus on a website build, and 3x as many tokens, because all that routing and stitching adds up. and it never clearly beat just using one good model on its own.

GLM 5.2 was the standout for value. nobody dethroned Opus.

Pricing's $20 / $100 / $200 a month (every tier includes both Fugu and Fugu Ultra), or pay-as-you-go Fugu Ultra at $5 in / $30 out per million tokens. no free tier, and it's not in the EU at launch.

Bottom line, it's more tokens and slower answers today. but it'll get better, and the idea is the interesting part here, not the performance.

not yet, anyway.

Anthropic just put Claude inside Slack as an actual teammate

This is one of the most exciting things Anthropic has shipped in a while.

It's called Claude Tag, and it drops Claude straight into Slack as a teammate you @mention. it remembers context across days, takes real actions through your connected tools, and in "ambient mode" it can even act on its own.

Here's how it works.

You @mention @Claude in a channel or thread, it breaks your task into stages, works through them using whatever tools & skills it's hooked up to, then replies right there in the thread with what it made.

The big shift is that it's multiplayer. one shared Claude per channel, so the whole team can see what it's doing and pick up where someone left off (a proper step up from the old 1:1 "Claude in Slack").

it also builds context over time so nobody has to re-explain themselves, and with permission it can even learn from other channels.

Ambient mode is defs the standout - this is where Claude stops waiting to be asked and takes initiative, flagging what it reckons you need, chasing loose ends, tagging the right people on its own.

The example Anthropic gave: it spots a "can't log in" email from a customer and alerts both the support and engineering channels itself, before anyone's even noticed.

For the bigger teams worried about it going haywire, there's proper control. admins decide which tools, channels and repos each "Claude identity" can touch, set spending caps (per org, per channel, per person), and manage skills centrally. 

Memory is walled off per channel, so sales never bleeds into engineering, and none of your conversations are used to train the models.

It plugs into Gmail, HubSpot, Airtable, GitHub, Drive and all the usual apps you connect to your normal claude.

Now the but. it's Team and Enterprise only, it runs on Opus 4.8, and it replaces the old Claude Slack app completely.

I can see this being really valuable for even a team of 1-5.

they can spin up separate Slack channels for different parts of the business, marketing, finance, content, and run an always-on assistant in each, especially with ambient mode and scheduled check-ins.

And you can just chat to @claude in a private slack dm if its just you, but you get to benefit from those more openclaw/ hermes style features.

Anthropic says about 65% of its own product team's code now comes from their internal version of this.

get started here

Perplexity gave its agent a real memory

Agents are stupidly capable now. the one thing they're still rubbish at though, is memory - actually remembering things over time.

Think about what a true second brain would do. a page on every person you deal with, your customers, your brand voice, the decisions you've made and why. so your AI starts every task already knowing your world, instead of from a blank slate every single time.

People have been DIY-ing this for a while with Obsidian and GBrain (an open-source memory system from Garry Tan, the President and CEO of Y Combinator).

and now Perplexity has shipped its own built-in version - called Brain.

The distinction that makes it click, especially if you're newer to this:

Most AI "memory" remembers YOU. your preferences, your tone, that you like bullet points. that you like dark mode over light mode, etc.

Brain remembers the WORK. what the agent did, what worked, what flopped, what you corrected, and then it rewrites its own playbook overnight. As Perplexity puts it: agent that remembers your preferences vs agent with experience.

Quick names untangle, because Perplexity does love a confusing name.

• Computer is their agent that goes off and does your tasks.

• Brain is the memory for said agent

On access, it's a research preview for Perplexity Max only ($200/mo)

Privacy's handled well. it's built only from your own activity, you can see, edit or delete anything it's learned, it's not used to train models, and incognito sessions are left out entirely.

Perplexity's own testing claims +25% answer accuracy and -13% cost on repeat tasks.

I think Claude will inevitably release something similar in the coming weeks.

Also this week...

• Higgsfield Supercomputer 2.0 → the big upgrade is that it's now a self-driving AI agent (NVIDIA-accelerated), pitched to run your marketing end-to-end.

💡 Builder's notes

Learn anything with the /teach skill

My favourite find this week, by a mile.

/teach is a skill from Matt Pocock that turns any folder on your computer into a personal tutor.

You open a fresh folder for whatever you want to learn, run /teach, give it a quick brief (you can even paste in a website link or a GitHub repo to learn from), and it builds you a little teaching project.

Interactive lessons you scroll through, it keeps track of what you've covered and where you're up to, then quizzes you at the end to make sure it actually stuck.

I used it this week to learn GBrain and really liked it. proper "learn by doing", rather than watching a video and forgetting it an hour later. → /teach by Matt Pocock (repo: github.com/mattpocock/skills)

Scheduled tasks: local vs cloud (I finally nailed my system)

Scheduling AI tasks has always been a bit iffy for me, and here's why.

A lot of today's agents are local, meaning they run on your actual computer (Claude Code, Codex, Perplexity's Computer, Gemini Antigravity).

that's great, because they can touch all your files. but most of us work off laptops, so the second you shut the lid for the day, every scheduled task stops dead.

The one upside of local: if a 5am task didn't run because your laptop was off, it just runs when you next turn it on. not the end of the world if you're on your machine all day anyway.

But…

This week I went deep on Claude's cloud routines and finally nailed it. cloud routines run on Anthropic's servers, so they fire no matter what, lid open or shut. (before this i’d just run local tasks on my always on mac mini - but I know not everyone has one)

My rule now is simple. anything timed but not critical stays local (unlimited runs). anything that MUST run on schedule no matter what, like clearing my email inbox every morning, becomes a cloud routine.

Worth knowing the run limits: Claude Pro gets 5 cloud runs a day, Max gets 15 a day, local is unlimited.

The one annoying thing with cloud routines is they have to run from a GitHub repo. So I pushed my whole .claude folder (all my skills and context) up as a repo, and connected Composio (the service that connects your agent to all its tools) so the routine can still reach everything. working great this far!

Helm, a little VS Code extension I vibe-coded

Still very much a work in progress, but I built a VS Code extension called Helm that lets you see all your skills, plugins, memory, MCPs and hooks in one place, instead of digging through config folders to find them.

Pretty neat, if I do say so myself.

🧰 Tools to try

• /competitor-analysis skill → drop in a competitor and it pulls apart their product, landing pages and socials for you. gold for sharpening your own positioning. feed the output straight into Claude when you're planning your next campaign.

• Orior → AI UGC and consistent AI-influencer creation, built for scaling e-com product promos.

• Free background library → 1,000+ free backgrounds for landing pages, product shots and branding. bookmark it for the next time you're building a page and need a clean backdrop.

• iMessage App API for agents → lets your agents send interactive apps right inside iMessage, not just plain text. early, but very cool if you're building anything customer-facing over text.

🥣 Brain food

• Pokémon Go, but for cats → someone vibe coded pokemon go for cats lol

• Modern Wisdom #1093, Jefferson Fisher → "How To Have The Hardest Conversations Of Your Life." one of the most practical pods I've heard in a while, on actually communicating better with partners, employees, teammates, anyone.

• Go deeper on Claude Tag → a full creator walkthrough of everything in the Claude Tag story above, in action. worth it if that one got you curious.

• WTF Is an AI Agent Loop? Genius or Hype? → what is loop engineering and should you be using it

• Horse kicks tree, farts on dogs, then runs away

I've met some incredible people out here at Cannes Lions so far, and even ended up on a yacht yesterday with Jeffrey Katzenberg (founder of DreamWorks), the CMO of Uber, the CMO of Kalshi, and a bunch of other brilliant people.

Crazy.

Thanks for tuning in :)