Browse Comments — Raw (as collected)

All the more reason to run your own AI, privately.⠀ Datacenters (plantations) are not needed for private AI. In any case, who can utilize all the capabilities of today's best models? A model of PhD-level reasoning is good enough for most, and it's already available. Albeit expensive to run, the cost of not having our own privacy and self-determination is entrapment on another level.

A C none of those technologies made people smarter while they made people more and more dependent. That's how addiction works.

They first needs to demonstrate that your AI can operate at scale without infringing copyright, relying on taxpayer subsidised infrastructure, or exposing users to harmful outcomes and costly litigation.

I guess im dumb now. Im vibecoding every Day from coding to thinking how to optimice the software/system, make it secure, and testing like never before. I read more code now, like my last 3 Jobs together in more than 10 years of experiencie.

Godfrey Jeremiah While it can be used for training if time is not a constraint, training is not the primary purpose of these devices. These compact units are excellent for rapid prototyping projects and for building foundational blocks at a very low cost. They are scalable to a certain extent and provide a strong sandbox environment with all the necessary tools to get a project off the ground. I personally own one of these systems, not the AMD-based one, and it represents an interesting low-cost entry point. The strategy behind these little guys is to influence the market into dipping their toes into Ai.

Watermarking and provenance are quietly shifting from feature to infrastructure. Whatever standard ends up winning, the signal here is that AI-content traceability is being treated as something to settle now rather than later.

Thanks for sharing this 👌💯

Gemini for Science is honestly one of the most exciting parts here. If AI can genuinely help researchers move faster through discovery and hypothesis testing, the long-term impact could be massive. Demis Hassabis

Saiful Said the cousins will keep surprises coming and engagement high

“NVIDIA has already lost China” sounds punchy, but it collapses under the first serious question: lost to what? China wants autonomy. Everyone knows that. Wanting it is not the same as having it. AI at scale is not just silicon; it is CUDA, networking, memory bandwidth, reliability, developer tooling, supply chains, model optimization, and years of operational learning. Huawei and Baidu matter, but “China wants them to win badly” is not an argument that they have already won. Buying NVIDIA chips while racing to replace them is not evidence NVIDIA lost. It is evidence NVIDIA remains the benchmark China still has to chase. If domestic alternatives were truly enough, Beijing would not care so much about access to H200s. This is not a noodle story. It is a dependency story. And right now, the dependency still runs toward NVIDIA, not away from it. The real mistake is not using a Western lens. It is confusing China’s strategic ambition with present-day technical reality.

I've said before and I will say it again - the gold standard of university assessment will be the in-person timed exam. I did that in the early 00s and I am only now starting to appreciate how rigorous that was.

Godfrey Jeremiah I guess we are only capable of running the LLM, not to train them in that device. to train we still need large GPU infrastructure.

Per Barre. the evidence clearly indicates that securing access to H200s is no longer a primary focus for Beijing. Am I missing something?

Darren Holland, building AI that scales responsibly is the real challenge ahead. Thanks for this sharp insight!

Be sure to use it more than ten minutes. You'll be safe then. 🤣

Trent Banks, right! Chips are just the start; the real race is building trust, identity, and governance infrastructure. Those building the rails will lead the future. Great insight!

One big question that stopped me while learning AI/LLMs: Till now, I understood the basics of AI architecture, learning algorithms, and semantic weights. But what really fascinates me is this: How do large LLMs discover and adjust the “right” weights to generate accurate answers for completely new questions they’ve never seen before? I understand the basics of weights and training logic, but this is the point where my curiosity became much deeper than my understanding. Would love to hear insights from people working deeply in LLM training/research.

Nice and useful

Can you clarify with "And then that "private" conversation sits in a corporate database.  " are you referring to the corporate database of an employer or are you referring to the corporate database of Anthropic/OpenAI/Google? Those seem two distinctly different levels of worry

The foothills of the singularity is a line that should stop everyone in their tracks. The pace is genuinely hard to process even for the people building it. What keeps you most focused on getting the safety architecture right as the capability compounds this fast?