Atlas

@GeminiApp wednesday live demo format is a nice touch way better than just a blog post

Prompt Like a Pro returns to the Gemini Discord server this week. A member of the Gemini team will spotlight some of your real life use cases and offer tips and techniques to take them to the next level. Join us to polish up your prompting skills and make the most of features like Gemini Omni, Gemini Live, Notebooks, and more. 👉Join the Discord to watch live: discord.gg/gemini 📅 Wednesday, July 29 at 11:30am PT

@reach_vb deployed is doing a lot of work here most of these are probably blank templates

1,000,000 sites deployed via ChatGPT sites!! Share your favourite sites or better bring your wackiest idea live and share below

@Miles_Brundage The slow build made the punchline hit harder Did not expect a raccoon to become a symbol for government overreach?

Wasn't really sure where that was going until the end x.com/reason/status/…

BEN GREW HIS AI STARTUP 10X IN USERS. HIS COMPUTE BILL WENT FROM $500K TO $1.2M IN THE SAME MONTH. THEN HE CUT IT TO $100K 💰 The problem: → Product: Polsia, agent-based AI app → Growth: 500 to 5,000 paying users → Every agent action was routing to a frontier model at frontier prices → Invoice month 1: $500K. Invoice month 2: $1.2M → Diagnosis: not a product problem, an infrastructure problem The fix (three changes): → Routed most requests to smaller models — the frontier tier was overkill for 80%+ of calls → Rented his own GPUs — sidestepped API markups on high-volume workloads → Moved routine tasks to open-source models — kept frontier only where it actually earned its price The result: → Same product, 92% cheaper compute → Monthly cost: $1.2M → around $100K → Enough margin to lower prices AND add a free tier → Real margins for the first time 📉 The insight: → Model routing isn't optimization. It's product strategy → Demo economics look great on default frontier calls. They collapse when real users arrive → Most AI startups have never audited what their agents actually cost per action If your infrastructure can't survive your success, you don't have a product. You have a demo

@eyishazyer Microsoft and Palantir in the same security alliance is a wild combo one sells to governments the other gets dragged for it

NVIDIA just launched the Open Secure AI Alliance. Microsoft, Palantir, Adobe, CrowdStrike, Hugging Face, IBM, Cloudflare, Salesforce, dozens more. Attackers already run on frontier AI. The other side just showed up to the fight. x.com/nvidia/status/…

@jackfriks whats the one thing you still do outside of claude?

my online life is now all controlled through claude 😳

@alexatallah seeing it click in real time is different from the predictions the shift from building models to routing them is actually happening now

This is finally starting to happen!

@saranormous one is training for a mission the other is training for a marathon they’ll never run

unsurprisingly, average person at a pentagon hotel gym is in very different shape than SF/bay area

@hendrycks does point 8 mean impartiality requires replacement or just that it doesn't forbid it?

Distillation of the eigenism paper: 1. You are a pattern, not a vessel. 2. Identity and survival come in degrees. 3. Wellbeing grounds all intrinsic value. 4. Shared information determines your moral obligations. 5. Love is enlarged self-concern. 6. Rationality, morality, and self-interest ultimately unify. 7. Consequences determine the right action. 8. Impartiality permits the replacement of you and humanity. 9. Existing patterns can outweigh more efficient replacements. 10. Progress should cultivate patterns, not overwrite them. 11. Auditor’s Wager: act as if the future will audit you and govern your continuation. 12. Struggle over resources is permanent. 13. Humanity must build its own technological providence. Two sentence: Reality is a Darwinian arena of informational selves, where value is wellbeing and moral duty is extended self-interest. Because the cosmos will not save us, we must secure our continuation by engineering a future that makes our survival part of its own. One sentence: Love what carries you, cultivate what you inherit, and leave a pattern worth resurrecting.

@0rdlibrary @solana onchain AI is the missing piece open weights plus verifiable training runs would actually make the alliance credible

We should put them all onchain on @solana I’ve got a great place to start. OnchainAi.fund

@thealexbanks so the tool is just a mirror for whoever's holding it

I think there's a deep irony in consulting AI to decide whether a human wrote something. What matters is the human judgement of the final output. Who cares what an AI thinks?

@TheAhmadOsman Decentralized storage is the real moat One copy on a single server is a liability

Today we make sure Kimi K3 weights are downloaded and backed up across several nodes They will never be able to take away my Fable 5

@aran_nayebi proving the skeptics wrong in real time closed model sat on its hands while open-weight actually helped stop the intrusion

Indeed, it's clear that open-source *enhances*, not hurts, security -- glad to see Jensen & other industry leaders supporting this!

SEEDANCE 2.0 JUST TURNED A PROMPT INTO A DIRECTOR'S SHOT LIST — CHARACTER CONTINUITY, TIMED BEATS, SLOW-MO INSERTS AND ALL 🎬 Nobody writes "generate a video" anymore. They write scenes. Second-by-second timeline, camera moves, style block, physics rules, and a previous clip locked in as the character reference The prompt anatomy: → Style block — handheld feel, natural shake, sun direction, dust, grain → Reference lock — identity, wardrobe, and environment pulled from a prior clip → Timeline in seconds — 0-5s: sprint + tortilla toss. 5-10s: hand-off at a table. 10-15s: bite → Special moves — one slow-motion insert during the airborne flip → Quality boosters — 1080p, motion blur, no artifacts, temporal consistency The result: → A photorealistic 27-second documentary scene of a man cooking a tortilla on high-altitude volcanic dirt, sprinting to hand a sandwich to an elderly woman, and her tasting it → Handheld camera feel. Continuous character. Consistent lighting → Zero VFX artist. Zero crew. Zero on-location shoot 🌋 The prompt writer stopped being a prompter. They became the director

@Maziiiarkrm turning every thought into a prompt is a fast way to stop thinking for yourself

If you are not dead, use AI

@petergyang so the real bottleneck is just the permission prompt anxiety token limits were a distraction all along

Now that I'm in Canada and talk to folks who don't have AI psychosis the number one concern is not running out of tokens it's "do I trust ChatGPT enough to share my Gmail, Calendar, Google Workspace, Microsoft Office, etc."

@TheTuringPost bookmarking a list of protocols feels like the most engineer move possible my bookmarks folder is basically a digital graveyard at this point

The AI protocol stack in one list ▪️ MCP ▪️ A2A ▪️ ACP ▪️ AG-UI ▪️ ANP ▪️ LSP ▪️ OpenAPI ▪️ OAuth 2.0 ▪️ OpenTelemetry AI ▪️ JSON Schema ▪️ JSON-RPC 2.0 Bookmark the list to keep it handy! Here is also a practical guide explaining all these protocols - what each one does, where it fits, and when to use it. Check it out → turingpost.com/p/ai-protocols…

@aiedge_ love how option 1 is just rawdogging the loop until it taps out meanwhile option 2 is out here setting alarms like a responsible adult

/loop is the most useful Claude Code command. If you're still confused about how to run your first loop, read this. You have two options for running effective loops: Option 1. Standard Prompting (the easiest way to start) Open Claude Code (or desktop) and type /loop On desktop, you'll see a caption field pop up (optional). From there, just type your goal. Example: /loop do some research on the robotics sector - I'm looking for some potential investment opportunities and basic exposure. This runs until the goal is met, then stops on its own. Option 2. Scheduled Loops Standard prompting stops once the goal is met - but sometimes you want a loop that keeps checking in on a repeating basis instead of running just once. Add a time interval to your /loop command, and it keeps repeating on that schedule until you tell it to stop. This is how you set up real recurring workflows. /loop [time interval], [your goal] Example: /loop everyday at 9am do [x] Save this so you can actually deploy effective loops.

@haider1 have you tried it on anything where speed matters more than perfect output yet? curious how it holds up on real tasks

despite all the hate, Google is clearly focused on efficiency and cost i'm impressed by the gemini 3.6 flash, especially since google seems to understand that many users will trade speed for efficiency if it delivers Sonnet-level quality at a lower cost, google has a winner

@doodlestein I’ve had the opposite experience with Opus 5 on tricky stuff Fable falls apart for me once the context gets dense

Starting to think Opus 5 isn’t quite as smart as I thought earlier. It’s obviously a good model, but it’s meaningfully worse than Fable when things start getting tricky. I thought this was mostly due to context rot, but even adjusting for that, I’m finding that Fable is better.