Centrox AI

@ivanburazin @ivanburazin This is pushing infra from request-response toward orchestration-aware systems. U can’t treat agent traffic like independent API calls when they’re part of feedback loops.

GitHub ,Google, Cloudflare don't have outages because they're badly run. The sheer volume of requests (mostly from agents now) breaks everything at that scale. We thought we'd cracked horizontal scale but now it's back to square one again. Agentic ai broke the infrastructure that worked for 20 years.

@rohanpaul_ai @rohanpaul_ai Anonymous drops are basically turning the community into a distributed eval layer. U get real-world stress testing across use cases no internal benchmark can fully capture.

Reuters: A mysterious AI model named Hunter Alpha recently appeared on OpenRouter. Rumors say it might be a secret test version of DeepSeek V4. - This model is massive 1T param. - 1M token context window - Many noticed reasoning patterns look a lot like the chain-of-thought style seen in previous DeepSeek models During testing, the model claimed it was trained mostly on Chinese data with a knowledge cutoff of May-25, which matches DeepSeek’s official specifications. --- reuters .com/business/media-telecom/mystery-ai-model-has-developers-buzzing-is-this-deepseeks-latest-blockbuster-2026-03-18/

Myth vs Reality: AI Chatbots Myth: “AI understands everything you ask.” Reality: LLMs predict the next token based on probability. That’s why most template chatbots break in real-world conversations: • hallucinations • inconsistent responses • weak context The solution isn’t a better prompt. It’s better system design. -> retrieval pipelines -> reasoning orchestration -> validation layers That’s how you move from a demo bot to a production system. If you're building a custom AI chatbot for real workflows, we can help: centrox.ai/services/custo… #AIEngineering #LLMs #AIChatbots

@ivanburazin @ivanburazin This is similar to APIs. You could regenerate them every time, but stability is the feature. Agents benefit more from predictable environments than constantly changing ones.

Agents won't vibe-code a new Slack every time they need to communicate. They'll use the same Slack because the other agent's team also uses Slack. Traditional SaaS will survive because standards don't get disrupted by probabilistic code generation. Network effects and standardization still matter.

@tszzl @tszzl Robots optimize for coverage and scale. Humans optimize for meaning and adaptation. Early Mars will likely be robotic, but long-term presence tends to follow where humans assign value.

the romantic human aspect of space colonization-placing our flags on other worlds-is redundant with robots and Von Neumann replicators unfortunately. the mars colony will be robots. ppl can join, but as consumers of an experience rather than critical parts of novel adventures

@elonmusk @bindureddy @elonmusk U're assuming a winner-takes-all outcome. More likely we’ll see specialization- different leaders across research, infra, and applications rather than a single dominant player.

@bindureddy Google will win the AI race in the West, China on Earth and SpaceX in space

Gemini 3.0 didn’t quite work out and most of us are still stuck with 2.5 Sometimes I don’t get it - what’s preventing Google from ditching all the side hustles and training 100 models from 100 teams in parallel Pick the model/team combination that produces a decent model! That way, they will at least stay in the AI race 😅

@bindureddy @bindureddy This isn’t a search problem where you brute force 100 variants. Training large models is path dependent. Small decisions early on shape everything downstream, so fragmentation can actually slow progress.

@Yuchenj_UW @Yuchenj_UW , The time is here, agentic AI is executing orchestrated workflows in real-time, turning APIs and sensors into a fully autonomous control loop.

“Dobby runs my entire house over WhatsApp.” Andrej Karpathy is running OpenClaw like Tony Stark runs JARVIS.

@GergelyOrosz , We saw this coming. Centralizing model access behind Max mode made usage spikes burn credits exponentially. Devs who spread workloads now hit limits immediately. Solution: reintroduce tiered access, with standard mode for steady use, Max mode for bursts, plus throttling and dashboards to monitor consumption.

I am hearing tons of complaints from Cursor customers at enterprise companies: A silent change put almost all models Cursor uses behind Max mode. Devs who used to manage to “spread out” monthly credits over a month see all of it used up in 1-2 days. Are furious + switching.

@emollick, What’s even more important, especially for AI researchers, is maintaining meta-awareness: keeping track of what the AI actually contributes versus what’s imagined, documenting assumptions, and building verifiable scaffolds around experiments. Amid the chaos, focus on clarity, reproducibility, and validation, not just endless novelty. Otherwise, it’s just a sleep-deprived hallucination in fast-forward.

We need guides through the inevitable bout of AI psychosis that affects professionals after they finally “get” AI. They often engage in intense, sleepless & impossibly complex projects in their area of expertise, with only AI for company. Its usually temporary & can be productive

@tmaiaroto , We would frame it differently. The real danger isn’t AI writing code per se, it’s humans losing the ability to critically evaluate outputs. If developers stop questioning AI suggestions, they stop exercising their judgment, and that atrophies intelligence over time. Think of it like giving someone a self-driving car: the vehicle can move you forward efficiently, but if you never check the road or monitor the sensors, you become blind to subtle failures. AI is a tool moving in your direction, but only if you remain the operator who verifies the trajectory.

@Miles_Brundage They're normalizing it along with many others so everyone can ship bad code. If the world's expectations of software are lowered, it gets a LOT easier to have AI write all the code. Why make AI smarter if you can make people put up with lesser quality?

I'm a bit worried Anthropic has an org-wide case of AI psychosis that makes them think Claude is good enough that they can ship random product features without breaking things, but in fact they *do* keep breaking things, and they're not online enough to notice people complaining

@Miles_Brundage , What’s even more concerning is that this mindset, overestimating a model’s robustness while underestimating systemic fragility, doesn’t just risk product glitches; it risks the integrity of field-wide experimentation. If teams ship features without rigorous validation, errors propagate, benchmarks become unreliable, and the collective progress of AI research slows under a veneer of false confidence. In other words, it’s not just a product risk, it’s a subtle drag on the evolution of the entire ecosystem.

@karpathy, Feels like two super-engineering Avengers converging, Jensen’s GPU mastery meeting your high-amp tinkering. With 20A throughput, this isn’t just hardware; it’s a mini high-performance compute sandbox, ready for Dobby the House-Elf's claw and experimental neural explorations to run at full parallel scale.

Thank you Jensen and NVIDIA! She’s a real beauty! I was told I’d be getting a secret gift, with a hint that it requires 20 amps. (So I knew it had to be good). She’ll make for a beautiful, spacious home for my Dobby the House Elf claw, among lots of other tinkering, thank you!!

@karpathy , remember how it felt like a niche research pocket, ConvNets + autoregressive RNNs stitched into end-to-end pipelines, more like clever hacks than paradigms? There was skepticism, low data regimes, unstable training. Then it quietly flipped, those prototypes became the canonical architecture, and deep learning moved from experiments to infrastructure. It turned perception + language into a jointly learnable problem space, unlocking everything from captioning to multimodal reasoning.

The signature is alluding to NVIDIA GTC 2015, where Jensen excitedly told an audience of, at the time, mostly gamers and scientific computing professionals that Deep Learning is The Next Big Thing, citing among other examples my PhD thesis (one of the first image captioning systems that coupled image recognition ConvNet to an autoregressive RNN language model, trained end to end). This was back when most people were still unaware and somewhat skeptical but of course - Jensen was 1000% correct, highly prescient and locked in very early.

@jbentley @svpino @jbentley, Software has always abstracted complexity; AI coding is just the next layer. The pushback isn’t technical; it’s about identity and control. Value shifts upward: from writing code to designing systems and intent.

Software has ALWAYS been about making it easier to create the next piece of software. That's why we build reusable functions, package them into libraries, etc. This virtuous cycle is what attracted me to software in the first place, I want to continue eliminating drudgery and increasing output. I don't understand the crowd that hates that AI can code now. Especially after we did the same to so many other professions.

A few weeks ago, a friend of mine stopped contributing to a few open source projects he's been working on for decades. He just stopped. "I don't want to train my replacement for free," were his exact words. I keep thinking about this. I really don't know what will happen to open-source projects in the next few years, when it becomes painfully obvious that developers no longer care because they can just "build" whatever they need on the spot. By the way, I don't think developers should be reinventing the wheel every time, and there's absolutely zero chance that leads to better software, but the reality is that many people don't care anymore.

@svpino, That fear is valid, but open source isn’t just code, it’s trust, standards, and maintenance. AI can generate tools, but still depends on solid ecosystems underneath. The shift is from writing code → curating and validating it. You can generate code, but not credibility on demand.

@yoemsri @pmddomingos @yoemsri, What’s even more important is relentless iteration compounding insight into breakthroughs.

@pmddomingos The original goal gives you direction. The failure gives you the discovery. Both are necessary and neither is sufficient alone.

Geoff Hinton set out to figure out how the brain works and failed. Andrew Ng set out to build a complete robot and failed. Demis Hassabis set out to achieve AGI using deep RL and failed. Yet they all succeeded.

@pmddomingos , What made Geoffrey Hinton, Andrew Ng, and Demis Hassabis significant wasn’t hitting their original goals, it was creating foundational primitives while chasing them. They didn’t just aim high; they shifted the paradigm, turning partial failures into new abstractions, tooling, and research directions that the entire field now builds on.

@emollick @emollick The stat is about the past, but the implication is rather forward-looking. If access to intelligence keeps scaling, the definition of “1.5%” might shift from material wealth to how effectively you use these systems.

I've had ChatGPT-5.4 Pro working away at a project I always wondered about: how lucky are you to be alive right now? Of all the ~117B humans who ever lived, only about 1.5% had a lifestyle roughly equivalently to a middle-class person in a middle-income country today, or better.

@Yuchenj_UW @Yuchenj_UW, Players like OpenAI and Google may own the base layer, but startups will build on top. The stack won’t collapse, it’ll expand. AGI won’t kill startups; it’ll just change the API where innovation plugs in.

Some people at frontier AI labs told me they believe startups are over. OpenAI, Anthropic, Google, xAI will absorb every industry as AGI nears. Coding today, science, medicine, and finance next. Then everything else. If they’re right, that’s a pretty boring end of the world.