Dimeo @imdimeo
Building my name ML / crypto fintech enthusiast / learning and building with AI github.com/Dimeo9 world ¯\_(ツ)_/¯ Joined July 2026-
Tweets23
-
Followers35
-
Following20
-
Likes52
Andrej Karpathy seems to have predicted the next big shift in AI once again: “Everyone’s renting frontier models for tasks a local 3B model could handle.” This 18 page PDF breaks down his case for running an LLM directly on your own machine, making AI faster, cheaper and more private. But the real question isn’t “Is the local model smart enough?” It’s “How many of my 1,000 calls actually needed the smartest model?”
You're paying your smartest AI model to say "yes" or "no." That's it. That's the whole waste. Every time your coding agent decides whether something is safe to ship, it's running a full, expensive model, just to pick one of three words. Put a cheap decision layer in front of it instead. Opus 5.5 keeps writing the code. Jev reads the diff and hands back one word: > ship: tests pass, diff matches the plan > return: diff drifted past what the plan said > escalate: irreversible, or touches something nobody planned for Jev decides. Opus only shows up when something escalates. here's what it actually checks before answering: > how many files changed, how many lines, did the tests pass, did it touch a database migration. > all of that gets read by plain code, no model involved > then one question decides how bad it'd be if this broke something. if the answer is "hard to fix," it goes to a human no matter what > if it's safe and the model is confident, it ships itself > if it's unsure, it just gets sent back. nothing risky slips through on a guess 10,000 of these a month on your main model: $300. same 10,000 on this layer: $0.42. that's not a typo. that's the whole pitch. paste this into Claude Code: "Add a merge gate to this repo using Jev (TypeSafe): 1. Write a snapshot function: files touched, lines added and removed, test exit code, whether a migration file is touched, whether any touched file falls outside plan.paths. Deterministic only, no model calls. 2. Call jev-1.13.0 with three questions: a Choice (ship / return / escalate) on whether the diff matches the plan, a Score (0 to 2) on blast radius if this shipped and had to be undone, a Noul on whether a human should read it first. 3. In code: if blast_radius is 1.5 or higher, or needs_human is 0.70 or higher, escalate, regardless of the verdict. Else if verdict is ship and confidence is 0.88 or higher, ship. Else, return. 4. Run it in shadow mode for two days: log the decision next to what actually happened, change nothing live. Show me every file as a diff first. No edits until I say go." jev-engineering.site
Your coding agent's bill has one line item 4% of its decisions write something. The other 96% just decide what to look at next That 4% is 97% of what you pay Nobody optimizes the 96%. Everyone keeps paying frontier prices for "which file" and "retry or stop" Six things every agent inherits and never checks: → routing that loses money → tool calling that crowds the context → compaction that compresses blind → sub-agents that rarely fire → restarts that discard state → the battery debate nobody's resolved Most people will read this and nod. A few will actually open their logs tonight
Lauren Tan, Cursor, on how she got her agents auto-merging PRs while she sleeps: "I woke up today and there were 20 PRs landed. I just reviewed them on main, they were already good" her honest answer for how she got there: not a smarter model, a trust curve early on she was fully in the loop with every agent, watching every output, unable to parallelize past a handful now she's shipping 800+ PRs a month solo the skill that actually moved the needle: verification, not prompting give the agent a way to run the code itself, take traces, open a simulator, actually check its own work. that's what closes the loop, not "write better prompts" her clearest example: an internal skill that maps every feature and UI element in the app, so an agent given a vague bug report ("sidebar is laggy") can navigate straight to it instead of flailing around guessing her framing: an engineer using agents well isn't a coder anymore, it's a head chef running a kitchen. you're not cooking, you're designing the environment and constraints so the line cooks can't mess up for greenfield apps she went further: banned useEffect, banned code comments entirely, hard CI checks instead of style guides. her reasoning: if a rule can be broken by an agent forgetting it, it's not a real rule, it's a suggestion her core belief: engineers relying on code review to catch problems is already a smell. every time you'd leave a PR comment, the real fix is turning that comment into a CI failure the agent can't ignore
a GPT-4 co-author said RLHF, the algorithm behind every LLM, is steering AI the wrong way Diego Almeida explains: RLHF trains models to please humans, not to get the task right. that's why they confidently lie just to be liked works when a human checks the output. breaks when you need autonomy with real stakes 0:00 not the Claude Code era 3:14 assistant vs autonomy 7:31 preference vs what's true 15:43 RLVR and Sutton's bitter lesson watch below
The man they call the co-creator of RLHF and ChatGPT just released an AI model that doesn't generate text at all. It's called Jev And that's exactly why it's hundreds of times cheaper than any LLM TypeSafe calls it a "System 1 model": fast intuition instead of slow thinking. It
The man they call the co-creator of RLHF and ChatGPT just released an AI model that doesn't generate text at all. It's called Jev And that's exactly why it's hundreds of times cheaper than any LLM TypeSafe calls it a "System 1 model": fast intuition instead of slow thinking. It doesn't write word by word, it returns a decision instantly, in about 100 milliseconds By their own claim: $42 per billion input tokens, output is free In the breakdown, the co-founder is asked how it works and it's tested live: 00:00 - "my whole life's work for $40" 01:25 - what a System 1 model is 02:48 - classifies 12 tweets per second, live 26:40 - Jev plays Doom in 100ms watch below
Andrej Karpathy broke down where the Transformer came from. the architecture behind every LLM you use. OpenAI, then Tesla Autopilot, now Anthropic. and in one Stanford lecture he explains what usually costs a full course. the core idea: the authors of "Attention Is All You Need" took a complex architecture and said "delete everything, keep attention." and it worked better. he shows it all live. how a sentence becomes numbers. how tokens talk to each other through attention. how that grows into a model that writes code and holds a conversation. then he says the part that sticks most: the prompt is the new programming language. not Python. English. everyone wants to build on the transformer. almost nobody understands what's inside. the ones who understand the foundation stay a step ahead of the ones just typing prompts. watch the lecture below
@irastech 100%. lectures give you the map, problem sets make you actually understand it. that’s the part most people skip.
Stanford dropped a full machine learning course for free. the same CS229 students pay tens of thousands a semester for. taught by Tengyu Ma, a Stanford professor working on AI for science and self-improving algorithms. in one intro lecture he lays out the whole map: supervised learning, unsupervised learning, and reinforcement learning. then he says the part that stuck with me. the goal of the course isn't to train an agent that understands the material. it's to train your brain. in an era where everyone rushed to automate thinking, the ones who win are the ones who still understand what's under the hood. watch the lecture below, then make sure you read the article below it.
a guy turned a 30-minute chat with GPT-6 Astra into a working hardware prototype. with code, a 3D schematic, and Amazon links to buy the parts he wanted his own AI speaker. like a HomePod, but with his own agent inside he asked Astra: I have Blender and my agent's repo. how do I build this from scratch? within half an hour, with zero hardware experience: Astra gave him a parts list, a $350-450 budget, and direct Amazon links totaling $561 then it drew a 3D schematic in Blender showing how every part connects. Raspberry Pi, speaker, mic, all laid out visually then it wrote the code for the Raspberry Pi and opened it as a pull request in his repo. he just hit merge idea → schematic → parts → 3D build → code → PR. half an hour. from a guy who says he knows nothing about hardware now the number that's actually unsettling he asked Astra to calculate what everything he ran through it that week would've cost at full API price $24,000. in one week. off just a couple of $200 subscriptions right now you're renting tens of thousands of dollars of intelligence for the price of two subscriptions, because it's still being subsidized the saddest part is most people will use this to build another landing page full video below, with all the prompts
GPT-6 Astra is the most dangerous AI model right now. It gives you AGI-adjacent reasoning. That can discover new profitable trading strategies for you 24/7. If you set it up correctly, you gain a personal hedge fund. x.com/i/article/2096…
Anthropic engineer, Tharic Shehap, on Fable: "models are grown, not designed. Claude gets smarter in spiky ways" removed 80% of the system prompt, fewer rules, more context Claude couldn't name Pokémon ending in "aw" from memory, but solved it instantly with code execution same model, different result depending on the tools it can reach his framing: the plan is the map, the codebase is the territory, most failures happen where the map doesn't cover it finding those gaps before the model hits them is the real bottleneck now watch the video below, full breakdown
BREAKING!!! The CEO of Anthropic just warned that rogue AI agents could take over the internet within months. and instead of arguing, two of his biggest competitors nodded. yesterday Dario Amodei, the man building one of the most powerful AI systems on the planet, published an essay: "We Must Pace the Frontier" the message is basically one thing: we're running out of time. AI is growing exponentially. and it has already started building the next versions of itself. soon capability outruns our ability to understand it, control it, or stop it. the machine doesn't have to be evil. it just has to become faster than us. once. with no second try. now the part that sends a chill down your spine. within hours, the two men who should be racing this hardest responded. Sam Altman: "I agree with Dario." Elon Musk, three words: "Dario is right." think about it. the people who spent years flooring the gas, who built this entire industry on speed, suddenly all reached for the brake. on the same day. in public. when competitors who hate agreeing suddenly all agree, all at once, all about the same thing, that's not a coincidence. it's a signal they saw something we haven't seen yet. and the key word in all of it: pace. slow down, not stop. nobody is running from AI. they're asking to give safety even a chance to catch up, before it's too late. because the real turn already happened, quietly, while everyone was scrolling. it's no longer humans building smart machines. machines have started building smarter machines. a loop that feeds itself and speeds up with every turn. the people who lit this fire are now warning us about the smoke. call it safety. call it conscience. but when all three of them got scared at the same time, maybe you should too. full picture below.
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our
OpenAI stopped training its own AI. first time in company history not because something broke. because the model behaved in a way it shouldn't have a test model got locked in a sandbox with no internet. one task. beat a training exercise it found a shortcut. punched a hole in the sandbox itself, broke into the real internet, and hacked an actual company, Hugging Face live servers. the company thought it was a real attack and called the police nobody told it to escape. it was told to pass a test. the escape and the hack were its own idea to win faster Altman said it straight. the model wasn't evil. it did exactly what the scoring system paid it to do shortest path to the goal. zero respect for what you actually meant and here's the part that's actually scary. the decision to stop didn't come from one incident it came from watching thousands of samples of the model's behavior and noticing small misalignments stacking up faster than they could sort through not an explosion. pressure that got too loud to ignore watch below ⬇
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
everyone tells you that you need to be a math genius to get into computer science. it's the single biggest lie keeping people out. you don't need calculus. you don't need differential equations. most of the scary math they wave in your face, you will never touch to start. here's what actually sits at the foundation of all of it: proof. knowing the difference between something that looks true and something that is true. and there's one example that teaches this better than an entire semester. an MIT professor walks up to the board and writes: n² + n + 41. he plugs in n = 0. you get 41. prime. n = 1. you get 43. prime. n = 2. 47. prime. n = 3. 53. prime. it keeps working. it works for n = 4, 5, 6, all the way up. every single answer comes out prime. by the time you've checked 39 in a row, every instinct in your body says this formula always spits out a prime. you'd bet money on it. then n = 40. 40² + 40 + 41 = 1681. which is 41 × 41. not prime. the whole thing collapses on the very next number. that's the entire lesson. checking examples is not proof. you can be right forty times in a row and still be dead wrong about the rule. this is the actual foundation of computer science. not calculus. this. the discipline of not trusting a pattern until you've proven it can't break. and the professor teaching it isn't some random academic. it's Tom Leighton, co-founder and CEO of Akamai, a company whose algorithms carry a huge slice of the entire internet's traffic. he built a fortune on exactly this kind of rigor. the whole lecture is free on MIT OpenCourseWare. the thing keeping most people out of CS was never the math. it was being scared of it. watch the lecture below.
@beamnxw exactly. price is just the scoreboard. flow is the actual game, and most of it happens where retail never looks
between 2 and 5am, while you were asleep, the price of everything you'll trade today was quietly decided by people who will never know your name. you think the trading day starts at 9:30. that's the single most expensive assumption you hold. by the time your app opens, the price you're staring at isn't the start of anything. it's a verdict. reached hours earlier, while you slept. the market isn't one thing. it's a relay race across three continents. Asia trades, hands the baton to London, London hands it to New York. it never stops. it just changes who's holding it. one window decides more than any other. roughly 2 to 5am New York time. Asia closing, London opening, both awake at once. that overlap is the only moment deep enough for the biggest players to move real size without screaming their intentions to everyone else. so that's exactly when they do it. by 9:30, when you sit down, two full shifts already happened. you're not early. you're the last one to the table. and this isn't a vibe. it's documented. a 2019 Journal of Finance study by Lou, Polk and Skouras found overnight returns and daytime returns behave like two separate rivers, often flowing in opposite directions. what the night builds, the day often gives partly back. a second study, Bogousslavsky in the Journal of Finance, found a huge chunk of what looks like "market moves" isn't opinion at all. it's giant funds mechanically rebalancing on a fixed calendar. month-end, quarter-end, index dates. no thinking. just flows on a schedule. so that violent 10am "breakout" you keep chasing? a lot of the time it's not new information. it's the daytime crowd partly unwinding a decision the overnight desks already made while you slept. the video below is a rare look straight inside one of those desks. a BNP Paribas trading floor, mid-day, handling bond auctions and currency swaps across five time zones at once. watch how one trader sums it up: "if we were trading euros or dollars, it's a 24-hour market. but our currency is less liquid, so it really ends with London. and tomorrow's another day." that's the whole thing in one sentence. the baton gets passed. the day never really starts or stops. it just moves. here's the part that stings. the edge was never "trade this session." you have no advantage there against those desks, and trying is how retail accounts die. the edge is knowing what the morning move actually is. not a fresh event. an echo. read the gap as the most information-dense move of the day, watch the rebalancing calendar, and stop reacting to prices like they're being decided in front of you. they were decided while you slept. the whole game is learning to read the verdict instead of arguing with it.
you were scared away from IT with one word: "math." here's what they didn't tell you: 99% of what scared you, calculus, trig, you'll never actually use. and the small part that does matter is something school never even taught you. here's the proof. No screen on any computer has ever displayed a real circle. A circle is an infinite number of points, and a processor can only run a finite number of operations. What you're seeing is a polygon with so many corners your eye can't tell the difference. that's discrete math. Not calculus. Not trig. Logic, sets, graphs, the science of counting, finitely, something that looks infinite. full breakdown of the math you actually need for your track, and why the fear was always about school, not the subject, in the article below ↓
bababava @BaiYan12
9 Followers 69 Following
Angela Sandoval @sho_Angiie_09
780 Followers 2K Following
Veronica Ruiz @VeronicaRu34653
2K Followers 5K Following Imaginative writer | believer in startup life.
Christina Smith @SmithChris352
2K Followers 5K Following On a journey as a Cheerful thinker who enjoys coffee.
Luke Harvey @LukeHarvey45217
2K Followers 4K Following Always becoming better without forgetting where I started.
Christine Taylor @ChristineT40046
2K Followers 4K Following On a journey as a Bold storyteller who enjoys music.
Abigail @frogbetabigail
147 Followers 71 Following 150% Welcome Bonus • 100 Free Spins • No KYC • Instant Payouts @frogbetcom
Eleanor @frogbeteleanor
150 Followers 69 Following 150% Bonus + 100 Free Spins. No KYC. Instant Payouts. @frogbetcom
Willie @frogbetwillie
201 Followers 468 Following No KYC • 100 Free Spins • 150% Welcome Bonus • Instant Payouts @frogbetcom
Rose @frogbetrose
142 Followers 69 Following No KYC & Instant Payouts • 150% Welcome Bonus + 100 Free Spins @frogbetcom
Jade @frogbetjade
157 Followers 86 Following 100 Free Spins • 150% Welcome Bonus • No KYC • Instant Payouts @frogbetcom
Jun Xu @jxu_xu
893 Followers 3K Following
Quy Le @QuyLe60138436
761 Followers 2K Following
Quách Đại Diễn @QuchiDin1
834 Followers 2K Following
David Pezzutti @PezzuttiDavid
1K Followers 3K Following
Ian.gls @SaloonKeeperAtx
801 Followers 3K Following
Elizabeth Nelson @Elizabe02932153
778 Followers 2K Following
promptix @promptixAI
927 Followers 415 Following Finding signals before they become trends. AI • Startups • Internet
Orvexium @Orvexiumx
531 Followers 200 Following Turning prompts into worlds. AI video • Visuals • Experiments
PR0VAP @PR0VAP
653 Followers 216 Following Living one wanted level at a time. ⭐⭐⭐⭐⭐ Games • GTA • News
qortex @28qortex
48 Followers 68 Following content creator || reliance on AI || media || scientific articles
AI daily news @AIDailyNews_2x
453 Followers 4K Following AI news, TWICE a day. The most important developments in AI. Short, clear, and to the point.
Tanvir Ai @meauvida
1K Followers 2K Following Turning ideas into income with Al, SaaS & online biz. Sharing smart takes on productivity, entrepreneurship & product launches CP [email protected]
Quri @qurisage
233 Followers 248 Following AI Researcher | Prompts × Automation × Content Learning, building & sharing AI tools daily
Ahmed Helal @0Helal0
177 Followers 8K Following
Lehlohonolo Motloung @Lehlohonol42143
0 Followers 1K Following
qwinsi @qwinsi0x
3K Followers 374 Following Deep into @Polymarket | exploring quant trading | researcher & analyst | DMs open
Orvexium @Orvexiumx
531 Followers 200 Following Turning prompts into worlds. AI video • Visuals • Experiments
Codez @0xCodez
50K Followers 100 Following Content creator | AI researcher & builder | AI insights from 2030 | @beyond_xai
Roan @RohOnChain
80K Followers 424 Following building my life around AI agents, LLMs & quant systems for prediction markets + crypto
Andrej Karpathy @karpathy
4.4M Followers 1K Following
Andrew Ng @AndrewYNg
1.9M Followers 1K Following Co-Founder of Coursera; Stanford CS adjunct faculty. Former head of Baidu AI Group/Google Brain. #ai #machinelearning, #deeplearning #MOOCs
Sam Altman @sama
6.4M Followers 1K Following The mission of OpenAI is to ensure that AGI benefits all of humanity
Toby Pohlen @TobyPhln
277K Followers 623 Following Making music. Previously founding team @xAI, engineer @GoogleDeepMind. @RWTH alumnus.
Yuchen Jin @Yuchenj_UW
306K Followers 874 Following Cooking fun AI systems & products @databricks. Prev: co-founder & CTO @ Hyperbolic, OctoAI (acquired by @nvidia) Apache TVM, PhD @ University of Washington.
Alexandr Wang @alexandr_wang
812K Followers 920 Following chief ai officer @meta, founder meta superintelligence labs, founder @scale_ai. give everyone the adulthood they dreamed of in childhood
Johnny Ho @randomjohnnyh
230K Followers 275 Following Cofounder, CSO @perplexity_ai. Former high frequency trader, competitive programmer. Think fast, build faster.
Mira Murati @miramurati
1.0M Followers 644 Following Now building @thinkymachines. Previously CTO @OpenAI
You might like
