AlgorithmicAmp

Father of two, Toronto. 1,500+ hours deep in AI—experimenting, building. Retail veteran. Books, debate, ideas. 44 & still figuring it out.

[AlgorithmicAmp @AlgorithmicAmp]

THE TRUTH ABOUT AI

WHAT'S ACTUALLY TRUE RIGHT NOW:

Synthetic data is real. AI models generate training data for other models. Happens at major labs - OpenAI, Anthropic, Google.

AI assists research. Tools help write code, analyze results, generate hypotheses. It's a powerful assistant for researchers.

Automation exists. Running experiments, processing datasets, computing metrics - parts of the pipeline are automated.

AI helps improve AI. Models evaluate other models, generate training examples, assist in design choices (e.g., search spaces, ablations) and training/eval settings.

We don't fully understand internals. Neural networks are partly opaque - and because of this, active oversight and safety evaluations exist.

WHAT'S NOT HAPPENING:

No "closed loop" exists. Humans set objectives and approve deployments. AI feedback is used for some steps (like RLAIF), but humans govern the process.

Humans make all strategic decisions. What to pursue, what safety tests to run, when to train, what to deploy.

No one's letting AI run wild. Are major labs such as OpenAI, Anthropic, Google, Microsoft planning to let AI recursively improve itself unsupervised? Absolutely not. That's science fiction.

THE ACTUAL CONCERN:

The real question isn't "are we letting AI run wild?" We're not.

It's "could we gradually lose oversight as systems get more complex, even while thinking we have control?" There's also a real technical risk: if AI trains only on AI-generated data without careful curation, quality degrades (model collapse).

That's a legitimate debate. It's about potential future loss of control, not current practice.

WHY THE CONFUSION:

"AI training AI" sounds like a runaway process. In reality it means:

A strong model generates coding problems to train a newer model
One model evaluates another's outputs
Automated hyperparameter searches
AI helps write training code

While humans aren't involved in every step of the process, they govern it - they engineer the systems, oversee the training, and build in checks and balances against critical errors.

[AlgorithmicAmp @AlgorithmicAmp]

@scaling01 They do make a lot of mistakes and give people delusions of grandeur. I'm currently on my second phase of AI generated grandeur.

[AlgorithmicAmp @AlgorithmicAmp]

@KevinSzabo14 Is 100 replies to posts what it actually takes? I had no idea it took so many and I only did about 10 yesterday. That explains a lot. Thank you for this.

[AlgorithmicAmp @AlgorithmicAmp]

@boneGPT I'd like to join if its not too late?

[AlgorithmicAmp @AlgorithmicAmp]

@AVGirl4Life Hi. I could use the help. Thank you.

[AlgorithmicAmp @AlgorithmicAmp]

I think this unnecessarily mystifies what are fundamentally engineered systems. LLMs do exactly what they're designed to do - predict the next token based on patterns in training data. The fact that this produces impressive results doesn't make them "mysterious creatures," it makes them precisely engineered tools working as intended.

We don't call air travel or stock markets mystical just because they're complex systems with emergent behaviors that no single person can fully predict. Complexity doesn't equal mystery. If these systems were truly unpredictable and all over the place, we'd dismiss them as useless. Instead, they work reliably because they're carefully designed human creations - just like every other sophisticated technology we've built.

When AI companies themselves use this kind of mystifying language about their own systems, it's not surprising that people interpret engineered complexity as something magical rather than what it actually is - sophisticated but comprehensible technology.

[AlgorithmicAmp @AlgorithmicAmp]

Why do the shittiest, lowest effort tweets go viral?

[AlgorithmicAmp @AlgorithmicAmp]

@codewithpri But reading and reviewing code is literally a core dev skill. If AI can write code that devs can successfully review and ship, isn't that developers relying on AI and is it too much?

[AlgorithmicAmp @AlgorithmicAmp]

@__Poisonivyyy Totally fair! And it seems to be working. You have 141.3k views as of now, so keep doing what you are doing!

[AlgorithmicAmp @AlgorithmicAmp]

@burkov Well yes and no. I do sometimes type in all caps and get mad at LLMs but I don't think it helps. lol.

[AlgorithmicAmp @AlgorithmicAmp]

Before you become an AI doomsayer, understand how LLMs actually work:

LLMs train on large mixes of licensed + publicly available data (docs, code, papers, forums, books) and synthetic data—with millions of new posts published daily.

They use pattern recognition and probabilistic modeling to generate responses. The process is more nuanced, but that's it in a nutshell.

One source going quiet? Impact depends on licensing and data access; big losses can dent specialized capabilities.

Much of the training data is out there, but copyright disputes are reshaping how it's accessed.

[AlgorithmicAmp @AlgorithmicAmp]

Mars by 2035 (speculative date) for SpaceX—what's your bet?
⭕ Yes, crewed landing
⭕ Yes, but later
⭕ Only uncrewed
⭕ Never
(Reply with your reasoning👇)

[AlgorithmicAmp @AlgorithmicAmp]

OpenAI designing custom AI chips with Broadcom - 10GW deployment

[OpenAI Newsroom @OpenAINewsroom]

We're partnering with Broadcom to deploy 10GW of chips designed by OpenAI.

Building our own hardware, in addition to our other partnerships, will help all of us meet the world’s growing demand for AI.

https://t.co/3vLZFPO0jF

[AlgorithmicAmp @AlgorithmicAmp]

Absolutely agree! I got a ton of likes and some followers from one reply a few days ago. Now my premium analytics graph has this one massive spike that makes everything else look tiny. And here I am trying to replicate it daily, chasing that next home run post. Spoiler: it's hard to hit home runs every day.

[Thomas Sanlis 🥐 @T_Zahil]

X payouts killed authenticity

Now everyone's grinding for that ONE viral tweet every 3 months while their regular content gets 5 likes.

We've all become engagement farmers, posting the same recycled garbage 🤦🏻

[AlgorithmicAmp @AlgorithmicAmp]

@JamesEbringer Fair strategy, but be aware: the health niche is a legal minefield. One wrong claim, one adverse reaction, one FDA letter—and you could face lawsuits that wipe you out permanently.

[AlgorithmicAmp @AlgorithmicAmp]

@apples_jimmy The last time I felt the warm embrace of love was when I kissed my children goodnight last night. Why can't I have both the warm embrace of love and an AI that I use simply as a tool?

[AlgorithmicAmp @AlgorithmicAmp]

Simple workaround for Claude's context limits:

At the start of your conversation, ask Claude to alert you when you reach ~30k tokens

Every 5-10 messages, ask: "How many tokens are left?" (insist on an estimate if Claude says it doesn't know)

Before hitting the limit, ask Claude to create a comprehensive summary of your entire conversation including all decisions, current state, and key context
Copy that summary and paste it into a fresh conversation with Claude to continue where you left off

Be proactive, not reactive. You control when to transition to a new conversation, not the system. No more getting cut off mid-project.

[AlgorithmicAmp @AlgorithmicAmp]

@VraserX I prefer the philosopher who overthinks, Claude. Also, Claude can write pretty long artifacts. I suppose that it's typical of a philosopher to rant anyway.

[AlgorithmicAmp @AlgorithmicAmp]

@cafreiman Hard disagree. Finding meaning in activism doesn't negate belief in the cause - it usually indicates deeper commitment. This framing just dismisses protestors without engaging with what they're actually saying.

[AlgorithmicAmp @AlgorithmicAmp]

Some days the algorithm loves you and it feels like you hit a home run. Other days it's radio silence and you're shouting into the void. Not complaining—just the reality of posting online. 🤷

[AlgorithmicAmp @AlgorithmicAmp]

From a behavioral perspective, this creates a paradox. Detachment from outcomes eliminates reinforcement contingencies - the very mechanisms that shape and maintain behavior. Without consequence-driven motivation, there's no basis for action. The framework contradicts basic behavioral principles.