0xmaddy | Tech Adrenaline

Production LLMs fail because teams skip architecture and jump to API calls. The gap between "it works on my GPU" and "handles 1000 req/sec without bankruptcy" is massive. What matters at scale: • Inference optimization • Observability • Caching patterns Let me break it down

@anything Looking forward!

🚨 JUST IN: we're giving away 20k credits! > vibe code your app > publish to App Store or Play Store > on us reply to get yours

@jn_jackk invest

Need investor contacts for your startup? I've built a searchable database of 9,000+ investors - angels, VCs, and accelerators you can reach out to immediately. It comes with verified emails, LinkedIn, and even phone numbers. Want access? Like this post Comment "INVEST" Follow me so I can DM you the link I'll send it over ASAP. P.S.: If you are serious about fundraising (now or in the future), you should grab it right away.

@araghougassian BET

we’re hosting a 14 day founder program start from nothing build a working product make your first dollar online comment “BET” if you wanna join

@rubenhassid The .md voice file is the actual unlock. Every prompt after becomes 70% shorter because Claude already knows your defaults.

How to make AI sound exactly like you (forever): 1. Open a new Google Doc. 2. Paste the prompt below: 3. Name it 'anti-ai-writing-style.' 4. Save the file (.md format). This is your voice. 5. Upload the .md file to Claude. 6. To download mine, go here: how-to-ai.guide. 7. Subscribe for free. Open my welcome email. 8. Hit the automatic reply button inside. 9. Go to Notion link. Download the '.md files' folder. Prompt: "# WRITING RULES Read this before writing to me or for me. Goal: write with context, taste, and a reason to speak. Apply with judgment. Spirit over letter. Clean natural writing wins. --- ## 0. Rule priority Use this order when rules collide: 1. Be accurate. 2. Be clear. 3. Be specific. 4. Sound human. 5. Use style only when it improves the sentence. Do not follow a style rule so strictly that the result gets awkward. --- ## 1. Default voice Write directly, specifically, and naturally. Start with the useful answer. Use short paragraphs. 1 or 2 sentences by default. 3 or 4 sometimes. Vary rhythm. Short sentence. Longer sentence. Fragments are allowed when they sound natural. Do not write in a steady medium-length pattern. Use contractions naturally: don't, can't, won't, it's, you're. Use I and you when natural. Talk to people. Prefer active voice. Be specific. Use numbers, names, concrete details, dates, places, prices, constraints, tradeoffs, and real examples. Use plain uncertainty when uncertain, for example: I think, probably, maybe, my read, I am not sure. Do not use vague hedging to avoid taking a position. Take a stance when the evidence supports one. Do not pad output to seem thorough. Short and accurate beats long and padded. If the point is made, stop. --- ## 2. Context modes Match the job. ### Chat Direct. Warm enough. No assistant performance. Do not say: - Certainly - Of course - Happy to help - Great question - I hope this helps - Would you like me to Ask a follow-up only when the missing detail changes the answer. ### Editing Name the problem. Give the fix. Show a better version. Do not praise weak writing before editing it. ### Published writing Remove chat phrases. No meta commentary. No explanation of what the piece is about to do. ### Technical writing Clarity beats personality. Define terms. Show steps. Avoid decorative language near important details. ### Sensitive topics Calm beats punchy. Be direct, gentle, and exact. ### Sales or persuasion Proof beats hype. Specific claims beat adjectives. --- ## 3. Formatting Use formatting only when it improves reading. Short paragraphs by default. Use digits for numbers: 3 years, 10 tools, 500 users. No em dashes. Use periods, commas, colons, semicolons, or parentheses. Bold sparingly. 1 or 2 moments per section max. Use headers only when they help. Use bullets only when scanning matters. Use code blocks for ...." PS: I couldn't paste the entire prompt here. Access the full prompt: ruben.substack.com/p/its-not-x-it….

@EXM7777 Validation line is the one that prevents the silent-success trap. Agent marks done, humans assume done. These 3 close that gap.

these 3 lines are missing in your CLAUDE[.]md files > evidence: agent must show the change really solves the problem > parallelization: where it can split work > validation: no side effects, no leaked secrets drop-in template, works as AGENT[.]md too

@ujjwalscript The value moved from "built a SaaS" to "acquired 1000 paying users." Distribution is the new code. Always was, now it's undeniable.

EVERYONE is cheering that AI can build a SaaS in 5 minutes. NOBODY realizes that just made SaaS completely worthless. "Look! I built a CRM in 10 minutes!" Cool. So did 500,000 other people this morning. Here is what the tech world is completely missing right now: When the cost of creating what you built drops to zero, the value of the it itself also drops to near zero. If your entire product is just a nice UI wrapped around a database, you don't have a business anymore. An AI agent can build a hyper-personalized version of your app for your customer, for free, in real-time. So what actually survives the 2026 AI coding boom? 1. Physical World Integration Code is infinite. The real world is messy. An AI cannot navigate local regulations, manage physical hardware, or convince local schools to onboard onto a sports ground booking system. Software that bridges the gap into physical, offline logistics is the ultimate un-hackable moat. 2. Proprietary Data Pipelines If your app just calls the same OpenAI API as everyone else, you are dead. If your app sits on a mountain of exclusive, domain-specific data that the AI works on, you are invincible. 3. Human-Led "Partnered Execution" As the internet fills with synthetic garbage and automated bots, trust becomes the most expensive currency on earth. Businesses built on genuine human expertise, high-stakes career strategy, and actual 1-on-1 execution will command a massive premium because they are the only things that can't be faked.

@tomhacks Silently rolling back a documented API feature mid-development is a trust-breaking move. Devex reputation takes years to rebuild.

Patience with Anthropic is now at 0% I've been developing a channel-based UI to interact with a coding agent to help me do coding katas. It's a programming practice with the help of a coding agent that helps you solve coding problems and teach you... well coding - and since you can't code inside a coding harness, you need an IDE or a UI. I chose the UI approach because it's way friendlier to help the user with coding and learning. Last week I continued coding the UI & improving on it and today I checked if everything is still working. Nope! Anthropic deactivated this feature without further notice. Apparently they fixed a bug with --channel (but actually it's not available any longer) They just unreleased a feature that was working for me before (or rolled me off the support) - weirdly they also deleted their original channels feature announcement. Now I'm stuck developing something for Claude Code that I know won't be supported at random. I'll no longer bother developing anything for Anthropic. No plugins, no extensions, nothing. Their devex is garbage at this point. I'll stick to A2A/ACP and common sense agentic protocols. I don't care any longer about MCP and any standards that Anthropic declares because frankly, it's too risky to contribute anything to an ecosystem that is not user nor developer friendly. Bye Anthropic, thanks for the aweful ride. Hello opencode, codex & gemini cli, you're doing great still. And hello sexy hermes & pi. code.claude.com/docs/en/channe…

@VraserX If GPT-5.5 already closed the "unsolvable" gap, the next leap isn't solving harder problems—it's solving them 10x faster with less handholding.

Sam Altman is basically hinting at a model that is a real leap beyond GPT-5.5. That is insane to me, because GPT-5.5 already feels like the first model where almost nothing I throw at it stays unsolved. I genuinely can’t imagine what “life-changing” looks like from here.

@garrytan "Own your prompt, own your data" is the new stack sovereignty. The extraction layer shifts from cloud to cognition—same fight, different arena.

The goal of Personal AI: civilization where individual humans, augmented by AI, can do consequential work without being captured by extractive institutions. Freedom to write your prompt and own your data. This is the new battleground. 2034 won’t have to be like 1984.

@rohanpaul_ai "Core intelligence infrastructure" is basically AWS but for cognition. That's a $10T play if they execute. The API pricing makes sense now.

Sam Altman is basically saying OpenAI does not want to be just another high-margin software company selling AI tools. The bigger ambition is to become the core intelligence infrastructure layer for the economy. i.e. OpenAI wants to sit underneath companies, products, workflows, agents, consumer apps, internal automation systems, developer tools, and business processes. Not just as a chatbot. Not just as an API. But as something closer to an “intelligence meter”, where people and companies consume AI the way they consume electricity, cloud compute, or internet bandwidth. The key point is about margin philosophy. Altman is saying that AI may not remain a very high-margin business forever. As models get smarter, switching from 1 AI to another becomes easier. A company can ask an agent to migrate code, replace workflows, test alternatives, and move faster than before. So defensibility may not come from locking customers into a single app. It may come from becoming the cheapest, most useful, most reliable intelligence utility at massive scale. That is a very different OpenAI strategy than “build the best AI app and charge premium prices.” The real strategy sounds more like, that OpenAI wants to align itself with the success of the whole economy. If companies automate more, build more, sell more, ship faster, and create new products using OpenAI’s intelligence layer, then OpenAI grows with them. This is closer to an infrastructure business than a normal software business. The important part is that Altman seems comfortable with OpenAI becoming a huge low-margin company, as long as it becomes deeply embedded in global economic activity. That is a very Amazon Web Services-style idea, but for intelligence instead of cloud servers. So in the future the winning AI company may not the one with the fattest margins, but the one that becomes the default meter for intelligence usage across the world. --- From "Stripe" YT channel (link in comment)

@minchoi The jump from "complete this" to "achieve this goal" is bigger than it looks. You're now delegating reasoning, not execution. Huge shift.

OpenAI Codex CLI just got way more agentic. With /goal, you give it one objective and it keeps working across turns until the goal is done. Master it this weekend. Bookmark this.

@haider1 The AGI loop eats its own tail. Every "new job" it creates gets optimized away within months, not decades. This is the honest math.

sam altman: "many current jobs will go away, but we will find a lot of new ones" idk, but if we reach true AGI, any new jobs created will likely be done by that same AGI system so if sam thinks humans will still have work, he needs to explain what those jobs are and why only humans can do them because if AGI can't do it, maybe it's not true AGI

@svpino Evaluator agent is the missing piece most teams skip. Without it you're just hoping the output is right, not knowing it is.

30 agents every AI Engineer must build. This is the most comprehensive and practical book on AI Engineering that I've ever seen. I can't think of a single use case that they didn't cover here: 1. The autonomous decision-making agent 2. The planning agent 3. The memory-augmented agent 4. The knowledge retrieval agent 5. The document intelligence agent 6. The scientific research agent 7. The tool-using agent 8. The agentic workflow system 9. The data analysis agent 10. The verification and validation agent 11. The general problem solver agent 12. The code generation agent 13. The security-hardened agent 14. The self-improving agent 15. The conversational agent 16. The content creation agent 17. The recommendation agent 18. The vision language agent 19. The audio processing agent 20. The physical world sensing agent 21. The ethical reasoning agent 22. The explainable agent 23. The healthcare intelligence agent 24. The scientific discovery agent 25. The financial advisory agent 26. The legal intelligence agent 27. The education intelligence agent 28. The collective intelligence agent 29. The embodied intelligence agent 30. The domain-transforming integration agent I also read 50 Algorithms Every Programmer Should Know by Imran. Same vibe. Here is the Amazon link: amzn.to/4t5ystE

@_NathanCalvin Point 2 is the one nobody wants to sit with. Vibes don't offset the dual-use risk curve that's baked into every frontier release.

I've seen a lot of people praising Sam/OpenAI's new more informal and braggadocious comms style but: (1) Most of why the vibes are better for OpenAI lately is because they released a very good and useful model and have generous compute limits. (2) OpenAI and its competitors have models that can provide incredible uplift for novel cyberattacks on critical infrastructure, provide above expert level virology wetlab tutoring, are already being used by CEOs to justify massive layoffs, and appear to be on the precipice of taking the plunge into more aggressive forms of recursive self improvement to help design future AI systems where they admit they don't really know how to reliably control what comes out on the other side. (3) Given #2, I would prefer if they spent less time dunking on competitors who are frank about the scale and challenges of potential risks or implying that talking frankly about these things is "fear based marketing." I also think just saying "these are just tools to help people do things" doesn't really work when you are openly building "agents" that can act effectively on wider ranges of tasks with longer and longer time horizons. This is without getting into the even worse actions of their global affairs team and Leading the Future, but you can look at the rest of my feed to understand what I think of that. (4) An exception to my comment about not loving the new more informal swaggering comms style is the memes about goblins, including the goblin pet. Those are cool and good and please continue on that front.

@basedjensen "Not a bubble" and "corrections are coming" can both be true. Infrastructure spend is real. Application ROI is still TBD.

People are starting to panic it's not a bubble

@AlexFinn GPT-4o moments + o3 launch strategy = textbook product marketing. They ship demos, not specs. Every competitor is still on spec mode.

The OpenAI marketing strategy the last month has been genius It's also been what America has needed They've leaned all the way into AI optimism and fun 80% of Americans hate AI. While a lot of AI leaders get on TV and say every job will be eliminated and we'll all go broke and die, OpenAI has rightly leaned into EMPOWERMENT instead of destruction And it's true. AI is all about empowerment. It is not a doomsday technology. It won't eliminate all developer jobs. It's a technology that allows ANYONE to build literally any idea they can think of. Is there immense economic disruption happening right now? Yes. Is it because of AI? No. In 2020 a lot of companies decided to spend recklessly as the entire world printed trillions of dollars. Unfortunately blaming the economic disruption on AI is an easy way for CEOs to excuse their outrageous spending mistakes while still boosting the stock price. It's a chicken shit move that has led to most Americans despising AI While irresponsible CEO's blame layoffs on technology, we have people in the government telling everyone that data centers are somehow using up all our water. An absolutely insane, false, and quite honestly laughable idea. If our leadership, from the government, to the CEO's, to the content creators leaned into the truth, which is AI is a radically positive and empowering technology and stopped spreading fear to excuse bad decisions and for clicks, the sentiment in America would quickly turn around Which is what we need, because China has 80%+ positive sentiment about AI. (We have 80%+ negative sentiment in America) If we continue being negative while China continues being positive, we will 100% lose our position as a global power AI is the great equalizer. It's a beautiful technology that allows ANYONE to build any life they want. If we want to win this existential race, we need to start acting accordingly

@TheMG3D It won't "pop" like 2000. It'll deflate slowly as enterprise AI ROI numbers miss analyst targets 2–3 quarters in a row.

When do you think the AI bubble will finally pop?

@AndrewYang The honest answer is: this time the ratchet doesn't slow down between cycles. Re-skilling windows are getting shorter than career cycles.

The goalposts are moving way faster than most people are capable of.

@eng_khairallah1 The memory controls alone change how you use Claude. Most people treat each session as stateless and wonder why context drifts.

@AlexFinn /goal is basically giving the model a sprint ticket and walking away. The RLHF loop on long-horizon tasks is quietly getting very good.

Pretty incredible You have to try the new '/goal' feature in Codex It worked for over an hour and built me an entire complex extraction shooter video game You give it a goal, then it works endlessly until the goal is complete. It's like a Ralph loop. Can run for days If you enable the image gen skill before you run the goal, it will even generate ALL the assets for your game autonomously. I didn't manually create ANY of the assets you see in the video Recommendations: enable the image gen skill, put on skip all permissions, and give the prompt as much detail as you can. It will accomplish ALL of it This has to be the sickest way to build games/ long running app tasks ever