Gave my buddy my proven, killer Facebook ad creative that has generated thousands of leads He dropped it in his FB account immmediately. Result? $15k in new MRR in 1 month! Most would say “like/comment/share” to whore the creative away for some virality. Me? "Equity or I share it on X." Levels to this! PS - almost any business can scale like crazy on FB. Comment 'share anyway', and I'll give you the sop doc... And if he doesnt come thru, maybe I'll share the creative too 🤣

Medical billing companies? Dead. Credentialing? Dead. Benefits verification. Prior Auths. All Dead. I just built a HIPAA compliant version of @bot Having it log in to our EHR and submitting claims. Takes advantage of custom modifiers and additional modification codes. All compliant. Consistent. Defensible.

If you are now using "landed" in every sentence. Congrats! You are AI zombified. If you get heart palpitations when your usage limit resets in four hours and you're nowhere near your desk. Congrats, you are AI zombified. If you've spent 14 hour days on a product you vibe coded that still doesn't work. Congrats, you are AI zombified.

Claude, ChatGPT and Grok are not fighting over who writes a better email. They're fighting over whose app your whole life runs inside. They want to be your everything because the web as we've used it is going the way of the Outlook distribution list. The internet moves off pages and into assistants, and the question stops being which site do I go to and becomes which assistant do I ask. Most people think they're just picking a chatbot. Look at what got connected this year. Your money. ChatGPT Personal Finance went live May 15 through Plaid, read access across more than 12,000 financial institutions. Your accounts and transactions, sitting in the same window where you ask everything else. It's on the Pro tier. Claude gets there through a connector called Era instead of building it in-house, and Perplexity beat both of them to it back in April. Your medical records. ChatGPT Health opened in the US on July 23. Apple Health, hospital records, labs, medications. OpenAI killed the separate health tab first, because among people who had access, more than 70% of health conversations were happening outside it anyway. Nobody wanted a health app. They wanted to ask the thing that was already open. Everything else. This is MCP, and it's the piece most business owners have never heard of. It's a standard plug. Any software that builds one can hand an assistant a controlled door into itself. Your CRM. Your calendar. Your books. Your ad accounts. Claude's directory is in the hundreds of connectors and climbing. And the browser, which is underrated. These agents don't run a stripped-down robot browser the way scrapers do. They drive a real browser with your real login, your real session, your real fingerprint. Which is why they walk straight past the bot detection that's been strangling ordinary automation for a decade. So what should you do? Sign up with Chatgpt, Claude etc. Ask it to do something for you that you didn't think it could do. Something annoying, practical and repetitive. You would be shocked what it can do. Stay curious across all of them. Claude, ChatGPT, Grok. They leapfrog constantly. Whoever was best in March is not automatically best today, so don't get loyal. For me? Ads has been the bane of my existence for decades. 10,000 hours in excel and google sheets for sure. Plus tech, software etc. Today? I'm barely touching it, and it's doing better than ever before. Daily audits of everything, reports in my chat in the AM. Fast pivots. Crazy insights (your creative fatigue starts day 6??). Automated ads and campaigns. Work that used to take hours. Now, that's not quite possible for most people, as the ads API and MCP's aren't straightforward. But more on that in another post.

Was looking at open source AI that is HIPAA compliant. Started where we do hosting for this project on @Azure ... They have 'some' models that seem to be under their BAA, but the really old ones. Went to @FireworksAI_HQ. Told me it would be a $50k commitment. Went to @baseten got a BAA in <1 day. Kudos to you guys.

If you're building a big, complex feature in Claude, the kind that would normally eat hours or days, stop hand-holding it. Paste this: "Use ultracode to fan out multiple subagents, ideally use lower intelligence agents where relevant by simplifying the process, and just use much, much more targeted DAG-style fan out / fan in to get this feature done."

My buddy does six figures a month on TikTok Live. A competitor started botting his streams to tank his sales. Now he wants the scummiest follower service on the internet to point right back at them. What's the worst one you've seen? The live-joiner bots proven to nuke an account. Name them. Asking for a friend.

OpenAI pushed GPT-5.6. Grok 4.5 landed, and Elon's already calling it "Opus-class." Both are "AGI," obviously. Same gooroos swore GLM-5.2 was the one last week, and Claude Fable the week before. Anything for the impressions. The part I keep not caring about: which team won. Open source, closed source, Claude, Codex, Chinese lab of the month. I love them all. I run all of them. What I won't do is torch thousands of dollars on a million output tokens to do work a cheap little model could've done in its sleep. Or sit for days waiting on a frontier model to hand me back garbage. Paying frontier prices to OCR a stack of invoices is like hiring a brain surgeon to open your mail. The whole game is right model, right task. Stop marrying a model. Match it to the job. Boring example nobody's posting about, because it's not on a leaderboard: Gemini 2.5 Flash Lite. $.10 a million in, $.40 out. Run your calls in a batch and it's $.05 and $.20. Going in, that's about half the price of standard DeepSeek V4 Flash. And it's blazing, north of 300 tokens a second. It reads your images and videos and OCRs your PDFs. Now DeepSeek V4 Flash. On the reasoning benchmarks it's the smarter model. Beats Flash Lite on the PhD-level stuff like GPQA. But I'm not asking a model to cure cancer. I'm asking it to read a PDF. And DeepSeek can't even see the PDF. Text only, and slower. So which one? Depends on the job. That's the whole point. Analyzing photos, videos, or PDFs, it's Gemini 2.5 Flash Lite, every time. Faster, cheaper going in, and it can actually see. Save DeepSeek's extra IQ points for the jobs that need them. Then run a hundred of them in parallel on Google's managed agents, for pennies. That's the move. Not chasing whoever got crowned this week. The right cheap model, pointed at the job in front of you. Try it on your competitor's video ads, or your next stack of PDFs. Flash Lite, managed agents, in parallel. See what a nickel actually buys.

Did not expect this AI surprise... Facebook comes swinging out of the fences with a model that tops the charts in... navigating computer browsing effortlessly and orchestrating other agents. Very excited to try it out.

I tried Fable 5 this weekend. It's genuinely awesome. I'm still not going to use it. And this is the part everyone is getting backwards right now. I spent a few hours having it completely revamp my SaaS product. GPT5.5 took 3 days, and still had mistakes. Fable did a great rebuild in hours. Great work! But the part that actually did the heavy lifting wasn't the model being "smart." It was the model running (DAG) agent swarms. Dozens of little agents, each handling one small slice, all at once. The part nobody selling you the new model will mention: you can do that RIGHT NOW. Today. You do not need the newest, most expensive brain on earth to get it. Drop Claude into ultracode mode. It starts releasing the swarms on its own. It fans your task out across a pile of subagents instead of slowly choking one giant conversation to death. And you don't even need Opus to run them. Sonnet 5 does the same thing. On agentic work it lands a few points under Opus (63 vs 69 on SWE-bench Pro) at roughly half the price. On 90% of what those agents actually do, you will never feel the six points. You'll feel the bill. We are all still over-using frontier models. That's the whole game most people miss. The mistake everyone makes: buy the smartest model, point it at everything, pay a genius to rename your variables. Do this instead. Unleash the swarm. Then match the model and the effort level to the job each agent has. Cheap and low-effort for the grunt work. Save the expensive brain for the two hard problems that actually earn it. I rebuilt a whole product this weekend doing exactly that. Not because the model was a genius. Because 30 cheap agents working in parallel beat one expensive one thinking out loud. Try it. Drop into ultracode, tell it to run swarms, set the model and effort per task, and let it fly. Then come tell me how it went.

Social media (kan)goorooa are all fear mongering about fable disappearing into expensive oblivion. Meanwhile I'm happily running sonnet 5 on ultra code... IYKYK

I love @Codie_Sanchez, but seriously?? The cheapest "employee" at my company costs $100 a month, runs on AI, and does the work of five people. Watch someone try to sell you a $19/hour assistant instead. There's a take making the rounds: a $19/hour virtual assistant is worth more than $4,000 in AI token spend. Cool. Two problems. 1) Who on earth is spending $4,000 a month on tokens to do VA work? For $19 an hour I can hire a physical human, in an office, IN CALIFORNIA. Nobody serious is torching four grand a month in tokens on admin. That number exists to make the VA look cheap. 2) A $20 to $100 a month AI subscription is a force multiplier on the person already in the seat. Not a line item to cut. Real example one of my businesses: My partner and one employee SHARE a single $100 a month Claude subscription. One of them does development, copywriting, AND creative. Three jobs. Five hires, the old way. One person now. The other does biz dev, ops AND ships updates to the custom backend of our lead-gen agency. Daily. Actual new features, not "I'll get to it next sprint." That is $50 to $100 a month. Total. So here is the honest test. Hand someone a $20 to $100 a month AI tool, and if you cannot find at least a 5% lift in what they produce, for every $1,200 a month VA you are paying, never mind a $3,000 one, stop blaming the tool. That is not an AI problem. That is operator error. The folks telling you a cheap VA beats AI usually have VAs to sell you. Follow the money.

'chinese models' are giving US companies a massive edge. @cursor_ai composer (prior generation?) is a fine-tune of Kimi k2.6 @ornith_ just released a fine-tune of qwen3.6 that beats opus4.8 Chinese models are forcing the market to not overcharge for insanely large models, and bringing a scrappy take on AI to the world. Thank them. I do.

Who spends $19 for a VA? Unless you're using @somewhere You could have an in person personal assistant for that. And how the hell do you spend $4,000 on AI and not see a return? Buy each employee a $20/mo @OpenAI chatgpt subscription. Let them connect to their email, project management system, anything they are using. Talk using it. Or do it with @claudeai desktop. Make it 'do stuff'. If they haven't doubled productivity, both of you need some serious training. If they are even slightly technical, get them the $100/mo Claude subscription. Or @CommandCodeAI Or antigravity by @GoogleAI . If they aren't custom coding and automating away your business. You've got the wrong people.

in 2026 there's zero excuse to keep saying "i wish someone would build me that app." just open claude code or codex. even if you still type with two fingers. heck, you don't even have to type anymore - just talk at it. speech to text does the rest. tell it what you want. watch it build the thing you've wanted for years. by the weekend. no dev team, no five-figure invoice. ahh, but you tried it. it looked so beautiful. and then it breaks. and the fliggerbidgeting begins. you paste the error back in. it "fixes" it. breaks two other screens. you spin in circles. nothing works. and you throw your hands up and walk away from the thing you were so excited about. that's the exact moment i built reframe for. it opens every screen of your app in a real browser and shows it back to you - desktop, mobile, tablet, side by side. flags what's broken. drops UI suggestions and plain-english comments. one click sends the whole list back to your AI to fix in one pass. no more copy-pasting errors one at a time. run it tonight. grab a free gemini api key, then just tell your AI to use github.com/resultkitchen/ref… -d and watch it fix your app lickety split. open source. comment REFRAME and i'll send the 2-min setup. 👇

Life hack to happy relationships is copying @mondaydotcom My wife is enamored with design. Good design/UX = good product. Prompting 'make it look like monday' makes my marriage happier.

Back to Opus it is...

What our good attorneys put on the bottom of their emails: *Please Note*: We bill additionally for reviewing/responding to client-submitted AI materials (and AI review of our materials). No advance estimate or approval will be requested. WDYT?

To my sorrow, Opus4.8 is nerfed. Returns more corporate-y judgemental non-answers than GPT5.4. Google is doing much better with flash-3.5, but still not quite there. Going to head back to codex, with it's retardedly low limits at the $20/mo plan. Depressing, like the industry took a step backwards this past week.