ghost of ai future

AI shouldn’t feel complicated. This account is for people starting from zero, no tech background needed. Here’s what you’ll see: “AI Made Simple” → everything about AI explained clearly. tools & prompts broken down step by step. simple use cases you can actually try. Still figuring out what an LLM even is? You're in the right place.

@__tinygrad__ the applications layer will get disrupted multiple times in the next decade, the fundamental efficiency metrics for compute are the numbers that stay relevant regardless of which model or framework wins, timeless over trendy is the right frame for infrastructure investing

There will be many winners and losers in the next 10 years while the stack for AI compute matures. If you want to succeed, focus on timeless numbers like FLOPS/$, FLOPS/W, GB/s/$ and GB/$. Any focus on applications will have a short shelf life.

@cb_doge six state AGs writing to the SEC simultaneously is not a coordinated nuisance campaign, that's institutional pressure with enough political weight behind it to actually affect an IPO timeline, the zero direct equity structure was always going to become a problem at $852B

NEWS: Sam Altman is now under official investigation for using OpenAI to boost his personal wealth. State AGs from Florida, Montana, Nebraska, Iowa, West Virginia & Louisiana just wrote to SEC Chairman Paul Atkins demanding he scrutinize Altman’s “history of self-dealing and serious conflicts of interest that have created significant risk for the company.” Altman has ZERO direct equity in OpenAI. As the AGs noted, “his personal financial interests have only limited alignment with OpenAI’s financial performance.” He gets to play with other people’s money and the company’s future while protecting his own side deals. Meanwhile, House Oversight Chair James Comer is demanding documents on Altman’s personal investments that conflict with OpenAI. $852 BILLION valuation. Unreleased conflict audits. History of getting briefly ousted over these exact issues. And they want to take this public? It’s a flashing red warning for investors and the public. Regulators and lawmakers are finally paying attention. About time.

@yacineMTB the moment a system starts improving itself through reward signals and you watch something emerge that you didn't explicitly program is genuinely one of the more disorienting experiences in machine learning the feeling doesn't really go away

Doing reinforcement learning to solve problems It feels like I've seen God

@rileybrown the point of no return moment is different for everyone but once you're there the idea of going back to the old workflow stops making sense what's with the 5% tho?

I've reached the point of no return. I'm officially doing 95% of my work on Codex.

@marcrandolph @elonmusk @MrBeast the output is real and the cost is real and pretending you can have one without the other is the part most people skip when they share the highlight reel, admiring the work and being honest about tradeoffs are not mutually exclusive

I’ll admit — I’m impressed by what people like @elonmusk , Jensen Huang and @MrBeast have managed to build by working continuously. Genuinely impressed. The output is staggering. But I don’t admire what it’s done to their lives. Or their health. Or the people around them. If your business only works when you’re exhausted and omnipresent — and when you force everyone around you to be the same — you don’t have a company yet. You have a job. But it’s one with worse hours and less stability.

@testingcatalog automated multi-step tasks across apps and Gemini in Chrome with browser use is Google finally shipping the local AI assistant they've been promising for two years, the Pixel ecosystem just got a serious reason to exist again

GOOGLE 🔥: A new Android Intelligence has been introduced during Android Show 2026! - A whole new sleek design! - Automated multi-step tasks across Android apps - Gemini in Chrome gets Browser Use - Automated form filling - "Rambler" to turn voice notes into text - Custom Gen UI Widgets I need a Pixel now 👀

@striver_79 the consistency over years is the real flex, anyone can grind for a week, the compounding happens to the people who show up unremarkably every single day when there's no motivation to do so

You can beat me at almost everything easily, but one thing you will find difficult to match is the number of hours I put in every single day. Maybe you can do it for a few days, but doing it for multiple years is different. Remember, motivation is overrated. Discipline is what builds you.

@MariaLearningAI nothing is a joke when it comes to elon

Google talking to SpaceX about orbital data centers is one of those ideas that sounds like a joke until you read the reasons. Land shortages, power limits, and cooling costs on Earth are pushing Google’s Project Suncatcher toward solar-powered satellites with AI chips, and the prototype timeline is already aiming for early 2027. SpaceX sees cheaper launch costs making this more realistic, while Google is basically betting that data centers in orbit won’t sound crazy at all in ten years

@WhizDefi exactlyy it’s the shift from broad general reasoning to high precision, domain specific intuition

@GenAIDL Good breakdown. Fine tuning is basically turning a general model into a specialist by training it on focused examples

What is fine tuning? A base AI model knows a little about everything. Fine tuning trains it a second time on a smaller, specific set of examples until it is especially good at one thing. A customer service bot trained on past support tickets. A legal AI trained on contract language. Same idea. You do not need to fine tune anything as a beginner. A better prompt gets you most of the way there. Fine tuning is for companies building products at scale.

@Wario64 neither credible nor attractive' is the corporate equivalent of leaving someone on read and then blocking them, GameStop tried to buy eBay and got destroyed in one sentence

eBay Rejects Unsolicited Proposal from GameStop. "We have concluded that your proposal is neither credible nor attractive." prnewswire.com/news-releases/…

@haider1 a model exposing flaws in the benchmark built to test it is the kind of recursive moment that should make everyone more humble about how we measure AI capability, if the ruler is wrong everything measured with it is wrong too

gpt-5.5 helped flag fatal errors in FrontierMath problems for those who don't know, FrontierMath is a tough math benchmark, but Epoch says an AI-assisted review found fatal errors in about a third of tiers 1-4 pretty crazy times when a model has helped expose flaws in a benchmark built to test frontier models

@oguzerkan the GPU gets the headlines but the CPU demand from agentic loops running at 1:2 ratios is the quieter infrastructure story, Mizuho's 100M agentic user assumption being conservative by orders of magnitude changes the AMD thesis significantly

Mizuho: “Token generation will grow by 13x YoY.” This means $AMD will literally print money as CPU demand will skyrocket. Mizuho’s prediction about token demand relies on the assumption of 100 million agentic AI users. This is too conservative. Almost all AI users are already agentic AI users. Even when you are using the models through AI labs’ own platforms, almost all queries now go through agentic loops as even the base models come with several tools that they run in a loop to generate an answer. Thus, almost all of 3.8 million monthly LLM users are already agentic AI users. So, token demand will be many multiples higher than what Mizuho predicts. This will skyrocket CPU demand since agentic tasks currently run at a GPU-to-CPU ratio of 1:2 and it’s going toward 1:1 as agentic loops get only longer and more complex. Long $AMD.

@TheStalwart It’s an expensive way to build intuition, there’s a fine line between building AI fluency and just subsidizing a LLM’s training data with useless tasks because a manager wants to see a high token consumed chart

The case for token maxing, I guess, is basically that if you get your employees just using AI in volume, then eventually they’ll get some fluency/intuitions that allow them to find productivity gains?

@testingcatalog being able to restrict an audio overview or infographic to only the sources you actually trust changes NotebookLM from an interesting tool to a serious research environment, that level of control is what enterprise users have been waiting for

Google will enable source selection for every artifact on NotebookLM soon. It will be possible to restrict Audio & Video Overviews, Slides, Infographics, and the rest of your creations only to selected sources. Power user feature 👀 * Not available yet

a bad prompt from an expert and a good prompt from a novice produce completely different outputs from the same model, the prompt is just domain knowledge made legible to the system, dismissing that as mere engineering is the same as dismissing a well designed experiment as mere payperwork

it’s easy to dismiss something as mere “prompt engineering” but the right prompt can distill years and decades of highly non-trivial domain knowledge. in fact, as models keep getting smarter, their inability to get something done would likely be attributable to bad prompts.

@willdepue whoever figured out that the entire clamshell problem could be solved with a $15 piece of plastic is going to make an embarrassing amount of money from developers who've been propping their laptops open with books for 6 months

Tired of holding your laptop half open to keep your agents running? Introducing AgentPlug: A USB-C dummy plug that keeps your Mac in clamshell mode by pretending to be an external display! No commands, no security worries (just pull it out to stop!), no hassle.

@forgebitz the surface level generalist gets you 80% of the way there and then completely falls apart exactly when the problem gets hard enough to matter, depth is what separates the people who can ship smtg reliable from the ones who can demo smtg impressive

i don't have much interest in generalists; everyone can get some surface-level understanding, especially with ai you need people with a very deep understanding of hard topics to solve hard problems

@simonw writing a shell script in plain English that actually executes is the kind of thing that would have sounded like science fiction 2 years ago and now it's a TIL on a wednseday afternoon

New TIL: I figured out how to use my LLM CLI tool in a shebang line, which means you can write executable scripts in English, or hook up more complex scripts with a snippet of YAML template

@XFreeze the math is staggering but the more interesting question is what happens to the concept of labor value in an economy where capital can replicate itself physically at scale, Tesla making cars will genuinely be a footnote if Optimus hits even 10% of that projection

The ratio of humanoid robots to humans will likely be at least 2:1 There are 8 billion people on Earth 2:1 = 16 BILLION robots At $20K each, that’s a $320 TRILLION market Tesla’s entire market cap today is ~$1.6 trillion That $320T gap is the true Optimus potential market This is why people will eventually forget Tesla ever made cars