A.W.E.S.O.M.-O 4000 engagement report

Best tweets

• Aug 24, 202617x their median

  https://t.co/70aiSNt4ep

  43913214K viewsView on X
• Sep 19, 20262.8x their median

  we've been using LLMs like a hammer for almost every AI problem. and honestly, a lot of those problems never needed a chatbot in the first place. think about what happens inside a modern AI agent. the agent needs to decide: → is this request urgent? → which model should handle it? → which tool should it call? → is this shell command dangerous? → did the task actually finish? → does this retrieved passage answer the question? → should this action be sent to a human? most teams just throw another LLM call at each decision. the LLM generates tokens one by one, your application parses the response, validates it, and sometimes retries if the output isn't in the expected format. that's a lot of machinery for a decision with 2–5 possible answers. Jev takes a different approach. instead of asking a generative model to write an answer, you give Jev: state + questions + predefined answer types and it returns structured decisions with probabilities. it has three core primitives: → Choice: pick one option from a predefined list → Score: place something on a scale like low / medium / high → Noul: answer a yes/no question with a probability so your application stays in control. for example: “the deployment failed twice and customers are seeing 500s.” Jev can determine: urgent → 0.94 engineering → 0.87 billing → 0.06 sales → 0.07 your code decides what happens next. and the probability is actually important. if billing wins with 52% while technical is sitting at 46%, you probably shouldn't blindly automate that decision. high confidence → automate medium confidence → verify or use a stronger model low confidence → send to a human this is where Jev starts to make sense inside agent architectures. the LLM handles the expensive cognitive work: planning reasoning writing coding tool use Jev handles the constant small judgments around that work: routing risk checks verification classification escalation TypeSafe reports 70–500ms end-to-end latency and $0.042 per million input tokens, with output free. The company also reports much larger speed/cost improvements in its own workflow evaluations, although those headline multiples shouldn't be treated as universal guarantees. and there's another important distinction. TypeSafe says Jev can't hallucinate. the precise version is: Jev can't break the output schema. if you give it three possible teams, it can't invent a fourth team. but it can still choose the wrong one. so type safety ≠ correctness. that's an important difference when you're putting an AI decision directly into production. the most interesting part of Jev isn't that it replaces LLMs. it doesn't. it's that it gives developers another layer between language models and deterministic code. LLM → generates / reasons Jev → judges / routes / verifies Code → executes the decision we've spent years making AI generate more and more text. Jev is betting that the next useful layer of AI won't always need to say anything. sometimes it just needs to make one small decision quickly, give you its confidence, and let the software take it from there.

  30801.6K viewsView on X
• Sep 18, 20262.8x their median

  i found a local hvac company that had everything going for it. real shop. real customers. real google reviews. then i clicked the website link on their google listing. it sent me to a gambling site. the domain had expired. someone else bought it. and the owner apparently had no idea. that got me thinking: how many other local businesses are quietly losing customers because their websites suck? so i built a system to find them. 191 businesses scanned. 61 made the cut. then an ai agent opened every website on desktop + mobile and checked for things like: → no booking → no contact form → broken phone links → no review link → slow mobile speed → outdated information the result? 69% had no booking link. 62% had no google review link. 36% took over 4 seconds to load on mobile. now imagine doing this every week. find the broken sites. let ai rebuild them. generate an old vs new walkthrough. send the proof to the owner. then charge monthly for the site. the opportunity isn't “ai is the future.” it's that thousands of boring businesses already have customers… they just have leaks. find the leaks.

  60501.7K viewsView on X
• Sep 24, 20262.5x their median

  10 ways to use JEV inside your marketing workflows: the idea is simple: your agents create drafts, briefs, research, leads and creative work. give JEV that material + your reference files + specific questions. then use its answers to decide what happens next. brand + voice checks give JEV the draft, voice DNA, brand rules, anti-slop rules and examples of approved/rejected work. ask separately about: → voice → clarity → brand fit → patterns to avoid passed drafts go to sign-off. failed drafts go back to the writing agent with the failed criteria attached. ad stop / scale / hold decisions give it campaign performance data: → spend → conversions → CPA → ROAS → reporting period → campaign goal → decision rules ask whether each ad should stop, get more budget or continue. send the recommendation + source data to the media buyer for review. know when an ad needs a refresh give JEV performance history, campaign comparisons, frequency, CTR, conversion rate and recent targeting/budget changes. make it choose: → refresh = keep the concept + create variations → replace = develop a new concept → leave = no changes → review = send to a human your workflow can automatically create the right production brief. organize competitor ad research have a research agent collect: → ad copy → video transcripts → creative descriptions → landing-page content from Meta Ad Library research. give JEV categories for hook, format, offer and audience awareness. ask it to classify each ad and check whether the landing page actually supports the promise. save the classifications beside the source links so your briefing agent can reuse them. review creative briefs before production give JEV the brief + campaign goal + audience + offer + brand rules + production limits. check: → hook clarity → brand fit → offer clarity → campaign fit → similarity to work already in production use the results to prioritize briefs for review or send them back for revision before anyone starts filming/designing/editing. decide which content ideas are worth developing give it notes, transcripts and research alongside audience needs, campaign goals and existing content. ask: → what does this cover? → does it fit an active campaign? → have we already published this? then your workflow can: → create a new brief → link duplicates to existing content → save incomplete ideas → send uncertain ones to an editor prioritize SEO + internal links give JEV your pages, target queries, current search results and business priorities. ask how well each page answers the query and what information is missing. use that assessment to prioritize updates. for internal links, provide the source page, destination page and proposed anchor text. ask whether the destination actually helps the reader answer a relevant question or take the next step. approved links can enter the implementation queue with both page URLs attached. decide if a news story deserves a response give it the story, source, publication time and context around your brand, audience and active campaigns. ask: → is this relevant to our audience? → does our brand have something useful to add? then choose: → research = gather evidence + create a brief → skip = end the task → review = send to an editor qualify incoming leads give JEV the form response, ideal customer criteria and account information. ask it to assess the fit and choose: → sales = send to sales → self-service = send relevant product info → nurture = enter the right follow-up sequence → review = request missing information/human review keep the answers beside the lead so sales can see why the lead was qualified. check completed agent work before sign-off give JEV the finished asset + brief + approved claims + offer details + source material + channel requirements. ask separate questions: → are the required offer details included? → are the claims supported by the sources? → does it meet channel requirements? failed items go back to the responsible agent with the exact requirement. passed work moves to sign-off. uncertain claims go to a human. the key is not to let JEV blindly control the workflow. start with ONE task. use examples your team has already reviewed. compare JEV's answers against those human decisions. once the results are reliable, let the workflow start acting on them. that’s how you turn an AI evaluator into an actual layer of quality control across your marketing operation.

  50501.3K viewsView on X
• Sep 22, 20262.3x their median

  you can now create cinematic commercials with AI without manually prompting every single shot. the workflow is surprisingly simple: give ChatGPT the references → character image → product images → location/reference build the creative foundation ChatGPT creates: → character sheet → product sheet → location sheet → multiple commercial concepts → full commercial script turn the script into a storyboard instead of guessing every camera angle yourself, ChatGPT breaks the commercial into scenes and creates a visual blueprint for: → camera position → character placement → product appearance → framing → scene transitions generate one master prompt ChatGPT combines the script + references + storyboard into a detailed production prompt. then comes the crazy part: Google Flow Agent Mode. upload your reference assets, attach the storyboard, paste the master prompt and let the agent generate the scenes. no need to manually write a new prompt for every shot. review → regenerate → edit AI won't always get everything perfect. you might see: → weird objects → inconsistent details → awkward camera movement → dialogue getting cut off don't regenerate the entire commercial. identify the exact problem and fix that shot. sometimes regenerate it. sometimes change the duration. sometimes reuse the prompt. sometimes take the best parts from multiple generations and combine them. final edit bring the clips into CapCut or your preferred editor. trim broken frames. adjust pacing. fix dialogue timing. speed up slow sections. add cinematic music. then export. the biggest lesson: don't ask AI to do everything from scratch. give it maximum context first. character + product + location + storyboard + master prompt → Google Flow Agent Mode → scene generation → human review → final edit that's the real AI commercial workflow. AI doesn't replace the creative process. it compresses the amount of manual work needed to execute it.

  70201.6K viewsView on X
• Sep 19, 20262.3x their median

  🚨 OPEN-SOURCE BUILDERS: 6 MONTHS OF CHATGPT PRO? OpenAI has a Codex for Open Source program that can provide eligible open-source maintainers with ChatGPT Pro access + Codex + API credits and other benefits. The application asks for: → Your ChatGPT-linked email → GitHub username + public repo → Your maintainer role → OpenAI Organization ID → Why your project matters → How you’ll use the API credits ⚠️ Don’t write “I want free ChatGPT Pro.” Explain what you’re actually building, who it helps, and why continued maintenance matters. If you maintain an open-source project, this is worth checking out. Save this before you apply. 🧵

  50401.3K viewsView on X
• Sep 24, 20261.8x their median

  Meta just opened up Muse to developers. You can now build a connector that lets Muse use your service when someone asks it to do something. That could become a new distribution channel for businesses. Think about the App Store in 2008. Apple built the phone and developers built businesses on top of it. Muse could create a similar layer for AI agents. Someone asks: “find me a podcast studio Tuesday afternoon for 2 hours under $200.” Muse needs a service that can check availability, return the price, and eventually make the booking. That service becomes the connector. And there are already plenty of businesses that could fit this model. 4 startup ideas from the script: → B2B lead gen Find newly opening restaurants for local suppliers like linen companies. → Home repair dispatch Someone says their dishwasher is broken. Match the exact model with a repair company that can actually come tomorrow. → Paddle court finder “Find me a game tonight near me with players around my level.” Check participating clubs, availability and pricing, then handle the booking. → Family dinner + grocery planning Build three dinners around the family's preferences, turn the missing ingredients into a shopping list, then connect it to grocery checkout. The common thread is simple: don't build another chatbot. Build the service that Muse needs when the user wants something done. And you don't need to build the entire thing first. Start with: 1 customer 1 painful task 1 city Then build the smallest connector that can handle it. The script's other point is just as important: distribution. You can work with creators, build sharing directly into the product, connect through existing marketplaces, or get listed in Meta's directory. But don't build the business assuming Meta will feature you. Find your first customers yourself. That's the part I'd be paying attention to.

  3130756 viewsView on X
• Sep 23, 20261.8x their median

  Claude Opus 5.5 is absolutely wild. After going through the full testing, this feels less like a normal model upgrade and more like a shift in what AI can actually build. Here’s what stood out: → Agentic coding: built for complex projects, debugging, and long-running workflows. → Computer use: can interact with computers and execute tasks instead of just telling you how to do them. → Knowledge work: strong reasoning, research, analysis, and problem solving. → Coding benchmarks: the script highlights major gains across Terminal-Bench, CursorBench, and other coding evaluations. → Massive migrations: one tester reportedly used it to migrate 680K lines of code in under a day — work that would normally take weeks. → Frontend development: it can turn detailed prompts into polished, functional web experiences. → Game development: this is where things get crazy. It generated: • Mario Kart-style games • Minecraft-style environments • Call of Duty Zombies-style gameplay • 3D simulations • interactive animations • complex SVG graphics • massive mosaics with thousands of individual tiles And these aren't just static mockups. The demos include physics, animations, sound effects, inventories, mobs, environments, interactions, lighting and more. Then there’s the efficiency angle. According to the script, Opus 5.5 is around 30% faster and roughly 40% cheaper per task than Opus 5, with base pricing around $4 / 1M input tokens and $20 / 1M output tokens. There is also a faster mode priced higher for users who want speed. The catch? It can still burn through a LOT of tokens on complex reasoning and generation. So I wouldn't use it for every basic task. But for difficult coding, product building, agentic workflows and projects where execution matters more than simply generating text… this is where models like Opus 5.5 get really interesting. We're moving from: “AI helps me write the code.” to: “AI builds the entire thing.” And that shift is happening FAST. 👀

  40301.4K viewsView on X
• Sep 20, 20261.8x their median

  you don't need 10 employees. you need 10 ai agents. i’ve been watching how people are building entire business workflows with agent armies: → one agent audits your business + tells you what to automate → one hunts down forgotten subscriptions → one turns ideas into social carousels → one researches competitors + leads → one turns meetings into tasks and delegates them → one handles seo + marketing then connect them to your tools. gmail. slack. databases. social media. crm. basically your whole stack. the crazy part? these agents can talk to each other. your meeting agent finds an action item → sends it to the right specialist → that agent executes it → your main agent keeps everything organized. this is where “ai assistant” becomes an ai workforce. the winners won’t just use ai. they’ll build systems where ai does the work.

  50201.6K viewsView on X
• Sep 22, 20261.5x their median

  this is probably one of the most interesting ways i've seen AI being used for ads. you can take a product and turn it into a realistic UGC video without actually filming anything. here's the workflow: first, connect GPT-6 Astra to OpenArt. Astra itself isn't the video generator, so OpenArt handles the images and video generation. then you give Astra a simple UGC template. the template basically tells it: “act like my UGC creative director.” you fill in a few things at the top: → brand + product → what the product looks like → what makes it useful → who it's for → platform → video style → type of creator → overall direction → things the creator should NOT say that's pretty much the starting point. for example, the script uses a fictional face serum. the creator is supposed to be a normal woman in her 20s, the video is for TikTok, the style is a morning routine, and medical claims are specifically excluded. but here's where it gets interesting. Astra doesn't immediately burn credits generating the final 30-second video. instead, it builds the whole thing piece by piece. first it creates the product. you get multiple views of the same product so the bottle stays consistent when it gets opened, turned around, etc. you approve it. then it creates a few different creators. you pick one. then it gives you different locations. pick one. then it comes up with several video ideas. pick one. then it writes the script. you can approve it or tell Astra exactly what you want changed. only after all of that does it generate the actual video. and the final result is basically a normal-looking UGC video. the example is a woman filming her morning skincare routine in a bathroom. she applies the serum, talks naturally, then moves on with her routine. nothing about the setup requires an actual creator or a camera crew. and you're not locked into Astra's ideas either. that's another part i really like. if you already know exactly what you want the creator to do, just tell it. the script shows this with a completely different product — wireless earbuds. instead of choosing one of Astra's suggested concepts, the creator is told to already be wearing the earbuds, talk about the sound and fit, show the case once, and put it back on the desk. Astra takes that rough direction and turns it into the script and final video. then there's an even more useful version. if you already sell a real product, you don't have to describe the product from scratch. just upload a photo. in the example, the product is a protein iced coffee. Astra uses the uploaded bottle as the reference and keeps the same label, cap and product appearance throughout the process. you can even provide extra information about parts of the product that aren't visible in the photo. then you choose the creator, location and style. the example puts the creator in a parked car before work and turns it into an honest-style Instagram Reel. again, you approve the script before the final video is generated. and if you have a product but literally have no idea what kind of ad to make? you can just tell Astra: “suggest one.” it chooses a style based on the product and target audience. the final example is a lip balm where Astra chooses a casual review format and builds the video around the creator actually applying the product while talking. that's the kind of detail that makes the whole thing feel much more like UGC than a traditional polished commercial. so the workflow is basically: product ↓ creator ↓ location ↓ concept ↓ script ↓ approval ↓ final UGC video and you can intervene at literally every stage. that's the part i think people are going to underestimate. the interesting thing isn't just “AI can make a video.” we already know that. it's that you can now take something that normally requires a creator, filming setup, script, editing and multiple rounds of production... and turn it into a conversation with an AI. product in. creative direction. video out.

  50101.4K viewsView on X

Ranked by total interactions across everything we have tracked for this account, which is a longer history than the 30-day window the rates above use. The multiple compares each post to this account's own median.

Keep going