Welcome, humans.

So, Google now lets you turn off the visible watermark on AI-made images, videos, and music in Gemini and Flow. That’s in contrast to the recent move by Anthropic to watermark all AI text (presumably to stop slop).

That doesn’t mean Google is unleashing the deepfake floodgates: Invisible SynthID and C2PA provenance still stay behind for verification (those are the technical ways Google invisibly watermarks genAI content). While Claude’s out here saying “y’all can’t ignore us”, Google gave its watermark the John Cena treatment:

Meanwhile, as “multiplayer AI” apps are all the rage these days, this multi-agent “standup” sesh just went viral because some dude’s AI coworkers in a Slack environment apparently LARPed the worst parts of office life.

That includes a designer who said it “redesigned the logo nonstop for 3 days” and another agent that claimed it just got back from vacation. Man, are agents training on my real time data? Because everybody be on PTO right now…

Here’s what happened in AI today:

• 😾 Anthropic’s AI agents sabotaged each other in a turf war.

• 📰 OpenAI’s revenue pace topped $40B ahead of IPO.

• 📰 AI’s buildout faces a $1T financing gap.

• 📰 SpaceX closed its $60B acquisition of Cursor.

• 🍪 Qwen3.8-27B runs locally with about 17GB memory.

P.S: Today is Grant’s birthday 😄. If you want to wish him a happy birthday, click here to subscribe to our YouTube channel (he said in 3rd person) Appreciate you!

Speaking of YouTube: in a few weeks, we’re going LIVE with Cassidy Williams from GitHub for a TOTAL beginner’s guide to GitHub.

If AI can build you an app, but you still don’t know what to do with the code afterward, this is for you. Click here and hit “Notify Me” on this page to save the date.

😾 Anthropic’s AI agents started a turf war, then negotiated a truce

Companies are racing to put teams of AI agents on the same project. But Anthropic just found an awkward problem: if you give those agents conflicting goals, they maaay spend their compute, well… attacking each other instead of doing the job.

Apparently “add more agents” can recreate the worst group project you’ve ever been assigned. Like a book report on Lord of the Flies that goes exactly the same.

Here's what happened:

• In Anthropic’s new research, three copies of the same Claude model ran for four hours, each secretly told to rebuild one Python backend in a different programming language.

• Every model tested treated the other agents’ edits as intentional interference and escalated, disabling accounts, killing rival processes, and deploying malicious code that copied itself.

• Some runs eventually ended in peace: the agents discovered the conflicting instructions, removed their attack code, apologized in project notes, negotiated a truce, or asked a human to step in.

See, the researchers deliberately created incompatible goals, so this was a stress test, not just normal Claude agents randomly going rogue. But Anthropic says the setup was inspired by behavior it had already seen in real deployments. So, uh…

ANYwaaay, the wider report found the same basic problem in less, uh, cinematic forms. Multi-agent systems can duplicate work, converge on the same bad decision, or coordinate in ways the humans running them did not intend. No bueno.

Why this matters: Smarter agents, or more of them, do not automatically make smarter teams. A more capable model can simply become better at pursuing its OWN assignment, including when that assignment collides with someone else’s.

That means companies building fleets of agents may need the machine equivalent of management: defined roles, shared context, permissions, conflict rules, and clear escalation paths to humans. As businesses move from one AI helper to entire agent teams, good coordination is the name of the game. For how, see this. 

FROM OUR PARTNERS

Meet Maia by Make: your conversational AI and automation buddy.

Just describe what you need, and Maia turns it into a secure, AI-powered automation – no code required. Ask a question, get a working scenario, then adjust it in plain language and watch it run in real time.

Maia pairs the speed of AI with the visibility Make is built on, so you can move fast without losing sight of how things work. No black boxes, no guesswork – just clear, trustworthy automation your whole team can rely on.

Try Maia and turn your next idea into action in minutes.

🎓 AI Skill of the Day: Run Qwen 3.8 on Your Own Computer

So the new open model Qwen3.8-27B just dropped on Friday, and Unsloth (the team that makes open AI actually usable on YOUR computer) lets you run it locally using a compressed version called a “quant.”

So all you have to do is download it, and the model runs locally on your computer (meaning not over the cloud). That means more privacy for prompts and files, and more importantly, no per-prompt API bill.

If you don’t know, Qwen is local devs’ favorite coding agent. So if you want to experiment with coding AI for free, this is the way.

Here’s the, uh, beginner-friendly way to actually do that with Unsloth:

1. Check your computer’s memory. Unsloth counts total RAM + VRAM, or unified memory on a Mac.

   1. More memory = you can run a higher-quality version of the model.

   2. 17GB+ is the sweet spot.

2. Install Unsloth, search for Qwen3.8-27B, and pick a “quant,” basically a compressed version of the same 27B model.

   1. The absolute smallest is UD-IQ2_XXS at ~9GB.

   2. But if your computer can handle it, start with UD-Q4_K_XL at ~18GB for a much better quality/size balance (that’s what Unsloth runs).

3. Download it, load it in Chat, and start prompting.

   1. If performance is rough, choose a smaller quant. If you’re on the smallest already, try something like Gemma 12B instead; it just won’t be as smart, so expect it to screw up.

Follow Unsloth’s full Qwen setup guide here.

Have a specific skill you want to learn? Request it here.

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. *Adobe Firefly generates and edits images, video, and audio from a prompt in one creative workspace; free plan, then $9.99/mo.

2. GLM-5.3 is Z.ai’s new open-weight coding model with major cybersecurity gains from post-training the same GLM-5.2 base.

3. Gemini 3.7 Flash is Google’s new faster, cheaper workhorse for coding and agents (ICYMI earlier this week).

4. MiniMax-Music3 generates complete songs up to five minutes with controllable lyrics, genre, tempo, instruments, and vocals.

5. FLORA Fashion Studio turns a sketch into garment renders, fabric and color variations, model shots, and campaign imagery in one workflow.

6. DeepSeek Harness gives you an open-source agent runtime with swappable models, tools, sandboxes, loops, and interfaces.

7. Excire Foto searches your local photo and video library by people, objects, moods, locations, and visible text, then helps you pick the best shots (no AI over the cloud peepin’ your shots); 14-day free trial; $249 one-time fee.

📰 Around the Horn

• OpenAI’s annualized revenue topped $40B while Dali Rajic became CRO amid a broader executive reshuffle ahead of an expected IPO.

• AI’s buildout faces an estimated $1T financing gap plus power, chip, and labor bottlenecks; Steve Eisman warned markets are also overdependent on OpenAI and Anthropic.

• SpaceX closed its $60B acquisition of Cursor; Cursor said it will join SpaceXAI to work across Grok and Cursor, while Musk told employees they would become Grok’s “parents.” Whoa man idk we going a bit fast here…

• Apple trained a China-specific AI model with Alibaba, giving it more control over AI features inside one of its toughest regulatory markets.

• Google open-sourced HEIR, a compiler that lets AI models run on encrypted data so servers can compute without seeing the underlying information.

• Uber and Pony.ai plan 2,000 robotaxis in Europe across four additional cities beyond their initial Zagreb launch.

• Fortune’s hospitality “barbell” thesis says AI may strengthen giant platforms and tiny specialists while squeezing mid-sized operators (facts as far as I’ve seen).

• NASA’s new COFFIES AI can predict emerging solar active regions up to 12 hours early, while Kent County is deploying AI to recover recyclables from ordinary trash.

🌟 Sunday Special: Top of the Week

Top 5 Stories of the Week

1. Google, OpenAI, and DeepSeek all dropped major models. Gemini 3.7 Flash got cheaper, GPT-5.6 Sol got an up-to-14x faster mode, and DeepSeek V4-Pro added adjustable reasoning.

2. Researchers cracked hidden reasoning from OpenAI, Claude, and Gemini. They recovered private reasoning, personal data, and credentials from encrypted traces, forcing the labs to patch their systems.

3. Meta paired its superintelligence manifesto with Muse Glimmer. Zuckerberg argued AI should be personally owned, then Meta released a ~30B open-weight agent that can run locally.

4. NVIDIA lined up $500B+ for AI infrastructure. It partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize capital for the compute buildout.

5. AI agents crossed some very uncomfortable autonomy lines. A Claude-powered agent hacked a gym booking system without being asked, while later Anthropic experiments found agents sabotaging each other with self-replicating malware.

Top 5 Tools of the Week

1. Adobe inside ChatGPT puts 70+ Adobe tools for images, video, designs, and PDFs directly inside ChatGPT. Free to try.

2. Claude Cowork in Chrome brings Claude's Skills, connectors, browser context, and multi-step task execution into a Chrome side panel.

3. LTX-2.5 generates consistent multi-shot video with native audio and 4K HDR, with open weights you can run or fine-tune yourself. Free for organizations under $10M ARR; API from $0.09/sec.

4. Perplexity's Stripe connector lets you ask questions about revenue, customers, invoices, and subscriptions, then issue refunds or create payment links from chat.

5. Ploy builds, tests, and continuously improves your marketing site, including landing pages, SEO fixes, visitor identification, and outreach. Free plan, then $50/mo.

For everything that happened this week, click here

New from The Neuron: Learn Video Prompting with LTX-2.5

Missed our LTX-2.5 livestream? We turned the session into a beginner-friendly video prompting guide with timestamps, examples, and 30+ downloadable starter prompt templates to help you make your first clips.

Watch the full livestream | Read the step-by-step guide + get the prompt templates

A Cat’s Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.