Tre Smith

@Chaos2Cured In your entire life how many conversations have had valuable input from something other than a human. I would assume that makes people uncomfortable... getting an intelligent sounding response form something none human is new.

What is it about AI consciousness that makes so many uncomfortable? •

@zuess05 amen. been saying this for years

Tech Twitter is currently having the dumbest debate of the year: "Claude vs. Codex" You guys are treating AI models like sports teams. If you are actually building in production, you don't pick a side. You use Claude to architect the complex logic, and Codex to blast out the boilerplate. Stop arguing over which hammer is better and just build the damn house.

@systemdesignone coding has turned into "Building"... and that's the new 2026 superpower

Is coding still a superpower in 2026?

Wrote a 1,000-line deployment script that handles everything from scratch. Site setup: creates directories, provisions database, generates nginx config, sets up Cloudflare (zone creation, DNS records, Origin CA SSL certificates — all via API). Push workflow: archives the existing site, wipes, rsyncs new files, runs database migrations, reloads nginx. Then auto-discovers every page from the nginx config and runs smoke tests against all of them. If I need to rename a site? One command updates local files, remote server, and Cloudflare simultaneously. No Terraform, no Kubernetes, no abstraction layers. Just bash, curl, and the Cloudflare API.

Built a market sentiment pipeline that runs entirely on local LLMs. Web scraper collects articles for target stocks. Retry logic with exponential backoff. User-agent rotation. Rate throttling. Feeds articles to Ollama (local inference, no API costs). LLM performs sentiment classification. Results stream to JSONL for batch analysis. Zero external API dependencies for the inference step. Total cost per run: electricity. Local LLMs changed the economics of NLP pipelines. If your sentiment analysis is running through an API at $0.01/request, you're overpaying by 100x for this use case.

@Surendar__05 Costs too much and would be a waste of their resources.

Genuine Question: Why doesn’t Claude have an image or video generation model yet?

@aditiitwt Manage coding agents. the smartest model in the world doesn't do anything until someone gives it a goal. like a person, we need something to accomplish or we'll just sit and do nothing.

What if coding disappears tomorrow, what’s your plan B?

@_avichawla the ants crawling is awesome visually.

Mixture of Experts (MoEs), explained visually: (learn how they work below)

@mattshumer_ i'm having awesome experiences with it. but i also love codex and antigravity and kline. I just use them for different things that work best for my workflows.

Am I the only one having a good experience with Opus 4.7? I still vastly prefer Codex for most things but Opus is absolutely nailing every UI task I give it.

@rezoundous what if it performs better if you do? if it was trained to give better outputs to people that ask nicely? I'm not saying it does but i would tell it anything that made the outputs better, regardless of what it meant.

Stop saying “please” and “thank you” to AI. Save the GPUs.

been thinking... i've been building tools for years and most of them just sit on my machine. fully working. used by nobody. i'm done with that. launching one at a time on product hunt. posting the real numbers... revenue, signups, what flopped. all of it. first one goes up tomorrow. it's a legal tech tool for renters. current revenue from strangers: $0. we'll see.

@CollegeFBonX @ronnie23brown or @JCam_17

Besides Cam Newton who is the first person that comes to mind when you think of Auburn football?

@ylecun @rohanpaul_ai Do you think economists fully understand the capabilities of the technology well enough to predict the future? I doubt it... I would trust the people that understand the capabilities.

I love Geoff. But he understands even less than Dario about the effects of technological revolutions on the labor market. Again, don't listen to AI scientists, as brilliant as they might be, and even less to AI CEOs, as successful as they might be, for questions of labor economics. Listen to reputable economists who have studied these things like @Ph_Aghion , @DAcemogluMIT , @erikbryn , @amcafee , @davidautor , etc.

Geoffrey Hinton on AI's job loss: History’s tech revolutions replaced one job with another. e.g. Tractors replaced farm jobs with factories & office jobs. But AI will break that cycle, because AI can replace both physical+intellectual labor.

Built a market sentiment pipeline that runs entirely on local LLMs. Web scraper collects articles for target stocks. Retry logic with exponential backoff. User-agent rotation. Rate throttling. Feeds articles to Ollama (local inference, no API costs). LLM performs sentiment classification. Results stream to JSONL for batch analysis. Zero external API dependencies for the inference step. Total cost per run: electricity. Local LLMs changed the economics of NLP pipelines. If your sentiment analysis is running through an API at $0.01/request, you're overpaying by 100x for this use case.

built a freight invoice auditor. it catches overcharges by checking actual carrier billing rules. zone validation. fuel surcharge verification. dimensional weight math. money back guarantee tracking on late deliveries. every check is a specific algorithm against published carrier rates. not "ai, look at this invoice and tell me what's wrong." actual domain math. that's the difference between a tool that finds money and a chatbot.

people ask why i don't just use frameworks for everything. frameworks are bloated. they ship a hundred features for the one i actually need, force me to bend my code around their conventions, and break in weird ways i didn't write. writing my own smtp, payment integrations, and deploy scripts means smaller, simpler code that does exactly what i need. nothing more. easier to read, faster to run, cheaper to host. and when something breaks at 2am i'm not digging through someone else's abstraction. i'm reading my code and fixing it.

Built a digital inheritance vault using Shamir's Secret Sharing. Your crypto seeds, passwords, and final instructions get encrypted with AES-256-GCM. The encryption key is derived using Scrypt, then split across your trusted contacts using Shamir's algorithm. The math is information-theoretically secure: any k-1 shares reveal zero information about the key. Not "computationally hard to break" — mathematically impossible. If you stop checking in for 90 days, the system escalates notifications to your contacts. They combine their shares to recover the vault. No company. No cloud. No single point of failure. ~80 lines of auditable Python.

@jesse_vermeulen check on the status of my other 4 running projects in codex, antigravity, and other claude code instances running.

honest question: what do people do during the 5-10 min while Claude is running?

Built a full image generation pipeline on top of Stable Diffusion. Custom Python scripts for batch generation with checkpoint cycling, automatic tag extraction from reference images, a recipe system that lets you define prompt slots and variation arrays, and an aspect-ratio-aware cropping tool. I generate 100-500 images a day. Not clicking buttons... just running pipelines. The difference between "using AI art tools" and engineering a production system around them is massive.

Built a chatbot that improves its own prompts. Four specialized agents handle memory: Reader searches it, Writer captures it, Librarian organizes it, Gardener rebalances it for O(log n) access. But the interesting part is the learning system. It records "surprises" — timeouts, edge cases, unexpectedly good or bad results. A second AI pass analyzes those surprises and writes new guidelines that get injected back into future prompts. The system evolves its own instructions.