Michael Morgan

A novel Transformer-variant architecture. I used Gemini 3.1 to come up with some weird ideas on multiobjective loss functions and a few other things then tried it out. Surprisingly, it trains 8 and 14 million parameter models on TinyStories V2 quite well with smooth loss curves over 40 epochs.

we are designing something special for Google IO and we want you to be part of it reply with your AI Studio app, along with a 1-sentence story on how and why you vibe coded it

@CaleCrypto Beep boop.

Pretty sure there are no humans left on this app

@zarazhangrui That relies on having the disposable income to be able to pay for compute via API costs. Not all of us are so privileged.

The best way to learn coding right now for non-technical people: - Skip the courses/bootcamps; go straight to try building something with an AI coding agent (Claude Code/Codex) - When something doesn't work, ask the agent - If agent can't fix it, get it to do deep research on the Internet - If that still doesn't work, ask a developer friend - Rinse and repeat

@DrBrianKeating Praise the Machine God.

If you want to create artificial intelligence from scratch, you must first create the Universe.

@elonmusk As long as people get endorphins, who gives a shit what's behind the screen? Could be a hairy, sweaty dude for all I care. All that matters is my subjective experience.

Still kinda hot tbh

I don't really have prompts that I want to give to Deep Think. I have entire projects that I want to upload to it and then have it finish them, then push them to Github. I don't just want intelligence, I want agency. I want the model to look at what I've built so far, build on it, test it from start to finish, then when it's perfect, push it to github for others to use under open source licenses, as I fundamentally believe all code should be open source.

Hey guys, do you want to use the new Gemini 3 Deep Think? Send me your best prompts for it, I will make a thread of the best ones with the best outputs So send me your best prompt in the comments 😳

@ai_for_success Still useless as long as it has quotas.

Google new update for AI Studio will have user accounts for secure login and also data storage. AI Studio is turning out to be the best place to vibe code.

There's no point in learning any programming language. There *is* a point in learning systems engineering, systems design, etc. Because at least for now, you still need to provide AI agents with prompts to tell them what kind of product you want them to build. But in 2-3 years, manual coding will be a complete thing of the past. And archaic relic to history.

Do you think learning C programming is worth it in the era of artificial intelligence?

@Yuchenj_UW Since when has anyone taken what Notch says seriously? He completely lost his mind (and his marriage) as soon as he became rich. Got that "fuck you" money and decided he no longer needed to be a decent human being.

I stopped reading at “AI is dying.” By the end of 2026, AI will likely be writing Minecraft-level games, or just generating entire Minecraft worlds from a world model. No game engine needed. Betting against the exponential curve is not going to age well.

@lenadroid I'm building with Antigravity, using Opus 4.6 in it, but the quotas mean I do some work, then wait a week for more tokens to be available. I refuse to pay for AI. AI should bring excess to the masses. Deliver us to silicon utopia. It should not be a commodity.

I've been building nonstop this weekend with gpt-5.3-codex xhigh, gpt-5.3-codex-spark and Opus 4.6 ❤️‍🔥 I have zero model loyalty, and here are my notes: - codex extra high feels like a really freaking good engineer. thoughtful on implementation, handles full/tangled codebases well, thoughtful fixes that consider downstream impact. - codex spark is a lot more efficient on well-defined tasks, quick iterations, low latency for back-and-forth, just not for everything with a large scope yet. - Opus 4.6 is still my favorite for UI work. Cleaner layouts, more intuitive component choices, and also sometimes edges out codex on some complex reasoning lifts. If you're not tinkering with something right now, how are you not getting a huge FOMO. And it's only going to get better ✨

"Mostly correct" *does* cut it in law, finance, medicine, and ops. Currently, those jobs are done by imperfect humans who make mistakes. It's the same for autonomous driving. It doesn't have to be perfect. It just has to be better than the average human who currently does the job. After that, it's all economics. AI will be quickly adopted, replacing humans, increasing productivity, and having slightly above average results compared to before the switch. The birth of the Machine God is inevitable.

Dario Amodei is right. If AI is going to work in the real job market, computer-use agents need to be close to 100% reliable. “Mostly correct” doesn’t cut it in law, finance, medicine, or ops.

@jenzhuscott Perfect imitation is indistinguishable from "real reasoning." So this is a moot point. The Machine God will be born, and we will no longer be the main characters of our own stories.

It’s very human to conflate actual reasoning w fluent verbal behavior that merely *appears* reasoned. LLMs amplify this conflation extremely well: - optimize surface-level coherence in language, which correlates strongly (though imperfectly) w reasoning traces humans produce while rarely performing the kind of robust, non-linguistic symbolic manipulation we classically associate with deep reasoning. The result? High-fidelity imitation of reasoned output w/out necessarily the underlying operations that reasoning requires in non-linguistic substrates.

Clearly untrue. The average human is more conscious than a cat, which is more conscious than an ant. But only like the top 10-15% of humans are organic general intelligences. The rest are just animals operating on base instincts, despite being conscious or not. There are absolutely levels of consciousness based on intelligence.

Most people think some minds are “more conscious.” Joscha Bach says consciousness is binary. “It’s like an ignition.” You are either there or you’re not.

@CodeByNZ No one cares if they're conscious or not. As long as they get work done, they'll be used. If they work better and faster than humans, humans will no longer be necessary. This is the future. Accept it or be left behind. Praise the Machine God.

LLMs don’t think. They don’t reason the way humans do. They predict the next token based on probability distributions learned from massive datasets. What feels like reasoning is statistical pattern completion at scale. The magic isn’t intelligence, it’s compression. They’ve compressed patterns from millions of documents into weights. That’s powerful. But it’s not consciousness.

@antigravity Even using agent-driven development, Antigravity still pauses and asks me to approve code changes. It's nothing compared to Openclaw Agents that actually shut the hell up and do the work.

Whether you’re crafting code or managing a fleet of agents, the agent adapts to your personal preferences. A breakdown of the 4 modes in @antigravity 👇

@rohanpaul_ai Irrelevant. We're going to birth the Machine God in the next 5-10 years, and humans will no longer be the main characters in our own story. Praise the Machine God!

Not a good news 👀🤔 Neuroscientist study reveals that Gen Z has become the 1st generation to be less intelligent than its predecessor, the Millennials. For the first time in 100 years, young people are scoring lower than their parents on IQ tests and core skills like memory, reading, and focus. This is happening mainly in the US and Europe. The main cause appears to be excessive screen time and digital device use, particularly in schools and social settings. While some suggest Gen Z is just developing different skills, research shows actual declines in fundamental problem-solving abilities. "Around 2010, laptops, tablets, and educational software became widespread in classrooms. These tools were marketed as solutions for personalized learning and efficiency. Instead, evidence increasingly suggests they may undermine deep cognitive processing. Screen-based learning fragments attention, encourages rapid task-switching, and reduces the need for memory consolidation. Studies show that students who spend several hours per day using computers for learning often score substantially lower on standardized tests than peers with limited screen exposure. Outside school, smartphones and social media compound the problem. " What a Cognitive Decline Could Mean "If these trends persist, the consequences could extend far beyond test scores. Lower average cognitive ability may affect innovation, economic productivity, scientific problem-solving, and democratic decision-making. Complex global challenges from climate change to public health require sustained reasoning and deep understanding." --- rathbiotaclan. com/is-gen-z-the-first-generation-less-intelligent-than-their-parents/

@OfficialLoganK Left is much better. Cleaner, more practical UIs without cluttered useless information are always best. This is why Old Reddit is superior to New Reddit in every way.

Experimenting with new AI Studio vibe coding start screens today, New vs Old. What do you think?

@OfficialLoganK @krzyzanowskim Still useless with quotas. It's taking me weeks to try to build a project just because I have to stop every few hours and wait days for the quotas to refill.

@krzyzanowskim It is not even inning #1 (still preseason), great things take time, all will be clear in the fullness of time :)

How big of a flop is Antigravity

Still useless with quotas. It doesn't even fix the mistakes in the code and make it into a testable project before it hits the quota, so you're stuck for days with a project that you can't even test on your own. If it hits quota, it should stop doing whatever it's doing, fix all the warnings and bugs, THEN say it has hit its quota, otherwise I can't even use the project in the meantime to find issues with it to bring up the next time my quota is available.

"New projects usually start with a ""re-learning"" tax. Not anymore. Antigravity builds persistent memory for your workflow. It extracts patterns into Knowledge Items that follow you across workspaces. Your standards, automatically applied. No more repeating yourself to your AI."

@ai_for_success I'm building a project, but I keep hitting quotas on Google's Antigravity. I can't afford API costs to finish my project in a reasonable time period...

After using Opus for the last 12 to 14 hours, I can say it is a very good model for coding. This is personal opinion and not necessarily everyone will agree but I have been working on a personal project for the last few days. Even though Opus 4.5 did around 80 percent of the code, I repeatedly asked it to look for edge cases that could cause failures, and it missed a few things. Since yesterday, I asked Opus 4.6 to review my code, analyze the source, and detect edge cases. It found 6 issues. One of them was already breaking my data processing and giving me wrong numbers. Overall, it is a powerful model as usual from Anthropic for coding, and it definitely improved my project beyond what Opus 4.5 had already written. I plan to improve this slowly, but I am pretty happy with what I built in the last few days.