David Wu

Just hanging out on Twitter reading and enjoying tweets, occasionally commenting.

[tonbi @tonbistudio]

Today's Hermes Agent Masterclass is the finale! Module 10 covers security, an important topic for running agents to ensure each agent can perform its tasks while minimizing exposure. Hermes has a ton of built-in security features. In this clip, I show how you can set approvals for each profile depending on your needs.

[David Wu @DavidWuX10]

@RealAlexJones Just put an AI generated tag on all real photos and no one will be able to determine if real or not

[David Wu @DavidWuX10]

@elonmusk $59.99 per month includes Starlink cellular for the Car

[Cybertruck @cybertruck]

Cybertruck FSD saves driver from head-on collision

[Ryan Saavedra @RyanSaavedra]

Mark Cuban (@mcuban) does a great job explaining how middlemen pharmacy benefit managers (PBMs) drive up the cost of prescription drugs for Americans, thus setting up the need for Congress to pass reforms

Cuban: The US has the highest drug prices in the world "because we are the only country that uses PBMs"

[David Wu @DavidWuX10]

@chuckcook I convinced my 73yo mother to get a model Y back in early V13. It was a little stressful teaching her. FSD v14 should be easy to show family.

[David Wu @DavidWuX10]

.@Tesla_AI I also do not like FSD going 71mph in a 55 in CHILL mode. Fix this @elonmusk

[Dirty Tesla @DirtyTesLa]

JuSt PuT iT iN sLoTh Or cHiLl

[David Wu @DavidWuX10]

@DirtyTesLa I just experienced this too. 71 mph in Chill near a routine speed trap area. Sloth dropped to 55mph forcing cars to zoom past me on both sides. I miss controlling the speed

[David Wu @DavidWuX10]

@DevinOlsenn Advanced users should be allowed to tweak settings while grandma gets 5 choices only @elonmusk

[David Wu @DavidWuX10]

@ryan_vangaalen @Tesla_AI @aelluswamy @DirtyTesLa @wholemars @AIDRIVR @chuckcook I live in CT and the leafs blowing caused V14.1.4 to slam on brakes from 50mph to dead stop multiple times. Now I cannot trust FSD when it’s windy in CT.

This is actually very dangerous. Worse than 13.2.9 driving into oncoming traffic to avoid Black tar repaired road cracks

[Elon Musk @elonmusk]

https://t.co/Mg0oHBVoxf

[David Wu @DavidWuX10]

@dillonloomis This was last year on 13.2.6-13.2.8 ish. Same issue. Happened every day on my drive. I would only take over if incoming traffic since I knew it was an ongoing glitch

[David Wu @DavidWuX10]

I use @grok to analyze my pool water based on it reading my test strips

Pool tests (7/10/25 AM): Liquid kit - pH ~7.2 (medium pink-orange), Cl >5 ppm (bright yellow, top of scale). Strips: pH 7.2, TA high ~150-200 ppm (green/blue-green), CYA ~100-150 ppm (pink), free Cl 1-3 ppm (underestimated post-shock), hardness 250-1000 ppm.
Full gallon shock boosted Cl to ~9-10 ppm—good for cloudiness. TA high despite pH OK; CYA rising from tablets.
12-hr Plan: Run filter 24/7. Retest pH/Cl/TA at 7 PM. If Cl <5 ppm, add 0.5 gal shock. Lower TA w/ 0.5 gal diluted muriatic acid (drop ~20-30 ppm), aerate if pH <7.2. Limit tablets; switch to liquid Cl. Brush if algae. Aim clear by Sat.

[David Wu @DavidWuX10]

.@grok told me there is no chance in a third party fixing the problem. It’s too entrenched on both sides. @grok thinks it will take an entire generation to fix this. Minimum 30-50 yrs.

[MOMof DataRepublican @data_republican]

@elonmusk A third party will destroy all the progress we have made. Please don't do this, @elonmusk !

[David Wu @DavidWuX10]

@elonmusk @grok I use @grok for important questions and I use @Gemini for website questions like what time a store opens

[David Wu @DavidWuX10]

These aren’t just opinions—they echo what researchers are finding. The tech world’s buzzing with this debate, and it’s not just nerds in labs. It’s about the future of how we know what we know.
Let’s break down the timeline again, step by step, to make it crystal clear. First, human text runs out, likely by twenty thirty-two, with eighty percent confidence. That’s based on solid data from Epoch AI, tracking how fast we’re burning through words. Second, AI starts training on its own outputs. Experiments show this leads to trouble within a few rounds—five to ten, let’s say. Each round takes one to two years, so we’re looking at five to ten years after twenty thirty-two. That lands us between twenty thirty-seven and twenty forty-two for collapse to hit hard, where AI can’t keep up anymore.
What does collapse look like? It’s not a single crash, like a computer shutting off. It’s more like a slow fade. Answers get vaguer, less creative. Facts start to blur—imagine AI saying a cat is a type of tree. It might still sound convincing, but it’s not useful. For businesses, researchers, or students relying on AI, this would be a big deal. It’s like a library where every book starts repeating the same few pages.
Could we avoid this? Maybe. If humans keep contributing, even a little, it’s like adding fresh water to a stream. We could also get better at sorting AI’s words, keeping the good stuff and tossing the junk. Some companies are already working on this, building tools to spot AI-generated text. Others are exploring new data sources, like private archives or global languages we haven’t tapped yet. These could buy us time, but they’re not a cure if humans stay silent.
Let’s wrap this up with a big-picture thought. This isn’t just about AI breaking. It’s about what makes knowledge worth having. Humans bring the chaos, the wonder, the mistakes that lead to breakthroughs. AI can help us sort it, share it, scale it. But if we hand over the pen, we risk a future where knowledge feels hollow. The timeline—twenty thirty-seven to twenty forty-two, with eighty percent certainty—is a warning, not a destiny. We can change it by staying in the game, by keeping our voices loud.
So, there you have it. If humans stop adding to the world’s knowledge, AI could hit a wall in about fifteen years, give or take. But with a little effort, a little human spark, we can keep the fire burning. Knowledge is ours to shape, and the future’s still unwritten

[David Wu @DavidWuX10]

Does asking @grok about LLM model collapse contribute to model collapse by 2040 with 80% certainty since I am posting what Grok thinks @elonmusk 🤖☠️🧵

[David Wu @DavidWuX10]

Why does this happen? Think of it like this. Human writing is messy, full of surprises, contradictions, and rare gems. That messiness is what makes it rich. AI, when it writes, tends to smooth things out, picking the most common patterns, the safest answers. If it keeps training on those smoothed-out words, it forgets the quirky, the unique, the stuff that makes knowledge deep. It’s like a band covering its own cover songs—each version gets flatter, less original. Scientists call this “error accumulation.” The AI amplifies its own mistakes, and soon, it’s lost the plot.
Let’s get specific about the timeline. If human words run out by twenty thirty-two, how long until collapse sets in? In those experiments, small models broke down after four to five rounds of self-training. Big models, like the ones powering today’s chatbots, might take a bit longer, but not much. Each round, or generation, is like a new version of the AI, released every one to two years. So, let’s do the math. If we hit twenty thirty-two and start relying on AI’s own words, five to ten generations could take us to twenty thirty-seven or twenty forty-two. That’s when, with eighty percent certainty, we’d see serious collapse—AI struggling to answer questions, mixing up facts, or just repeating itself like a broken record.
But is it really that simple? Not quite. There’s a lot that could change this timeline. For one, not all AI words are bad. Some researchers think we can mix human and AI texts carefully, like blending fresh and canned ingredients in a recipe. A study from twenty twenty-four, mentioned on Wikipedia, suggests that if we keep some human data in the mix, collapse might be delayed. Others are working on tricks, like tweaking how AI learns, to keep it from getting too stuck on its own patterns. These are like guardrails, but they’re not perfect. Without new human ideas, the road still leads to trouble.
Let’s zoom out for a moment. What does this mean for knowledge? If AI starts collapsing, it’s not just a tech problem—it’s a crisis for how we learn, share, and grow. Right now, knowledge is a dance between humans and machines. We write, AI organizes and amplifies. But if humans step out, and AI takes over, we risk a world where knowledge feels like a rerun. Imagine opening a book, only to find it’s a remix of yesterday’s news, with no fresh spark. That’s the shadow of collapse.
There’s another layer here, something deeper. Philosophers might ask: what is knowledge if it’s just AI talking to itself? Knowledge, at its heart, comes from human experience—our struggles, our discoveries, our questions. AI can mimic that, but it doesn’t live it. If we let AI become the only storyteller, we might lose the human touch that makes knowledge meaningful. It’s like replacing a campfire tale with a robot’s recitation. It’s not the same.
Now, let’s talk about what could stop this. Humans are the key. If we keep writing, researching, and sharing, we refill that pool of words. Even a little fresh input can go a long way, like a splash of water in a dry garden. There’s also hope in new tech. Some scientists are exploring ways to make AI learn from more than just text—think images, videos, or even real-world sensors. Others are building systems that never stop learning, adapting as they go, like a student who keeps asking questions. These could stretch the timeline, maybe even push collapse out past twenty fifty.
But we can’t ignore the risks. Some folks on platforms like X, where people share quick thoughts, are already worried. In twenty twenty-five, one user posted, “AI’s heading for collapse if it keeps eating its own tail.” Another said, “Without humans in the loop, AI’s just a feedback loop, degrading fast.”.

[Try TLGen App | Brandon Starship 🇺🇸 @BStarr119]

Cyber welds hit different 📐

@cybertruck is jet engine mechanic approved 💪