Unitree is already robot producer, they don’t need car manufacturing sector like Tesla in order to dominate.

It’s funny that whenever I am critical of Tesla, I get attacked by Tesla fans for being another nobody who questions the genius of Elon Musk, and whenever I write something that confirms something Tesla is doing, I get attacked for being a Musk fanboy.

Case in point:
https://bdtechtalks.com/2020/07/29/self-driving-tesla-car-deep-learning/

You’re absolutely right. LLMs will enable many more people to express their thoughts without having a lot of experience in writing. In fact, since I’ve written this article, LLMs have advanced even more impressively and are helping me in writing drafts and sorting my ideas. But the human element of writing will never go away.

True, but this can be “fixed” with just a few minutes of editing.

The problem with LLM is that it can write more thoughtful, well-structured and engaging pieces than 99% of writers today. You can say it has nothing unique to say, but who can come up with *really* novel thoughts these days?

What’s more, even if it can’t bring its own experience to the table, it can draw on the experience of anyone who has been worth to writing about in the past. How many people can do that?

In science, THERE IS NO “GAME OF LIFE”!!

Instead, there are the unstoppable forces of Entropy (basically, the Second Law Of Thermodynamics) and those forces (known to scientists) that are capable of temporarily neutralizing the forces of Entropy: All of these forces are governed by rigid scientific equations. The equations for “data”, “information” & “knowledge” fit tightly into our current overarching Standard Model Of Physics (chiefly defined by Shannon’s Entropy Law) — as reported in our latest SKMRI Knowledge-Physics Lab field-notes on this topic:
https://www.linkedin.com/pulse/7-easy-things-you-can-complete-today-protect-your-stevenson-perez-mhfkc/?trackingId=Br39InHSEWRjcpwV35sgIg%3D%3D

Unfortunately, the original AI-system inventors in the U.S. did not know these physics equations: Our AI-systems will become rapidly safer & more-reliable & more-trustworthy, as soon as we stop playing games — and start fixing these AI-system design-flaws.

Sincerely –
Your SKMRI Knowledge-Physics Lab Colleagues

Thanks!

I think it’s https://arxiv.org/html/2403.05440v1

It is really impressive, but it doesn’t sound like Zinsser. He would have never used words like “drafting” and “researching” (he hated turning every noun into a verb) and he certainly would not have used “streamline.” I feel like the model just reflected the content of the article, albeit in a very impressive way.

The rise of artificial intelligence in writing prompts a reevaluation of our craft. These tools, with their ability to generate text with a command, challenge us to consider what distinguishes the human touch in writing. It’s not merely the construction of sentences or the adherence to grammar rules that define our art; it’s the soul behind the words, the unique perspectives and lived experiences that each writer brings to the page.

To those navigating this new terrain, I offer this advice:

Embrace technology as a tool, not a cradle. Let LLMs assist you in drafting, researching, and editing, but remember that the core of writing—the insight, the nuance, the voice—cannot be outsourced to an algorithm. Use these tools to streamline your process, but let the heart of your work be distinctly your own.

Cultivate your voice. In an era where AI can mimic styles and generate content, the authenticity and uniqueness of your voice are paramount. Your voice is what will distinguish your writing in a sea of automated content. It’s your thoughts, your humor, your empathy that readers seek. Nurture it, refine it, and let it resonate through your work.

Focus on the human experience. AI may capture patterns and mimic styles, but it cannot live life. It cannot breathe in the autumn air, feel the sting of loss, or revel in the joy of a shared laughter. Write about these experiences, these immutable facets of the human condition, for they remain the domain where only humans can tread.

Engage with your audience. The future of writing is not just in the hands of those who write for the masses but also those who connect with specific communities. Find your niche, engage with your readers, and build relationships that transcend the digital divide. In doing so, you ensure that your writing does more than inform or entertain; it connects, resonates, and endures.

Continue to learn and adapt. The landscape of writing is ever-changing, and what AI brings to our world is just another chapter in the long history of our craft. Stay curious, experiment with new tools, and always be willing to adapt your methods while holding fast to the principles that make writing an art.

In this new era, let us not lose sight of the fact that writing, at its core, is an expression of our humanity. No advancement in technology can replicate the depth of human thought, the complexity of our emotions, or the subtlety of our interactions. As we navigate the future, let our writing reflect the richness of our experiences, the clarity of our insights, and the warmth of our connections with one another. In doing so, we honor the legacy of those who have shaped our understanding of the craft, and we contribute to a world enriched by the power of the written word.

Its inability to have depth of understanding and use concise language makes it yawn inducing.

*I’m taking this from the description pseudo-science because it will provide answers that aren’t even related to what was asked of the student.

That is what I meant. It enables you to continue your conversation with the LLM past the context window, though as you said, it sticks to length of the context window (e.g., 4k tokens). That’s what the article says too if you read it carefully.

Your explanation is inaccurate. It does not change the context window in any way. If the LLM has a 4k context window, it can only respond using the context of the latest 4k tokens. StreamingLLM makes LLMs more efficient by removing the need the reset the cache and improves accuracy vs LLMs that aren’t resetting their cache. It doesn’t make it so that an LLM with a 4k context window can accurately respond to a 128k token prompt. This article is spreading misinformation. Read the FAQ section here. https://github.com/mit-han-lab/streaming-llm

Hi Jonathan. StreamingLLM does not change the architecture of the model to expand the context window. What it does is shift the context window while maintaining the accuracy and the reused part of the KV cache. So basically, you can extend the conversation with the LLM into millions of tokens as if its context window was unlimited, but without making any changes to the model or retraining it. I hope this helps.

When hosting your own opoen source LLM

The representations don’t arise implicitly. In fact, calling them representations in the AI tradition is fanciful. They are not interpretable. They don’t deal in concepts. And the property of flexibility is not well-defined beyond, possibly, some engineering concerns. This opinion is in the same mold as the belief that Eliza was sentient. Just a bit more sophisticated, that’s all.

There is no doubt that LLMs will be useful. But they will, without any doubt, become yesterday’s technology of tomorrow very soon. People will start viewing them as commodities, and the excitement will shift to whatever new and unknown engineering marvels will appear.

And all that is excellent news. Long live technological progress.

But this angle does not have much to do with science. Understanding human cognitive functioning will require much more than fMRI experiments, the metaphor of neural networks, the success of the algorithm to select the most probable next word in a text and the starry-eyed pronouncements that LLMs are edging closer to sentience.

To some people, this problem is of interest as such, irrespective of its complexity and the very low probability of it being solved within their lifetime.

The problem is that nobody knows how to build “true” artificial intelligence. It also likely won’t happen in one shot. People do easy things before complicated things, and build upon prior work. We have figured out how to do image classification, voice recognition, art generation, seamless language generation. Next, we will have better imitators of human thinking, better software assistants, better robots. Then can afford to work on the harder problems.

Yes, sniping at snipers is always fun. 🙂

Relativity theory was a paradigm shift, yes, but it did not come in a vacuum. People by then had thoroughly explored Newtonian mechanics and electromagnetics. Those did great, but had problems in extreme cases, such as a fast-moving frame of reference and close to Mercury. Special relativity dealt with the first of these, and only later general relativity handled the latter. It also needed mature math machinery that developed at the same time.

The issues you raise, about versatility of the human mind and being able to do metacognition were known for decades. Unfortunately, there is no easy path forward. I think new methods will necessarily build upon existing methods, at least being motivated by failures in existing methods. So it will be incremental work either way, and there’s likely still plenty to do to understand how far current methods go.

Thanks for your comment. I think that it is fair to say that it is easier (and more amusing) to snipe than it is to create solutions, and easier still and even more amusing to snipe at the supposed snipers. But I disagree about a few things. The range of problems that can actually be solved is not a lot greater than it was. Large language models solve one problem, predicting words. People have been more clever in using that capability to address (usually with difficulty in practice) apparently different problems. Second, an incremental approach is not a sure thing. Relativity theory was not incremental relative to the physics that came before. Thomas Kuhn describes paradigm shifts that were central to scientific progress. Finally, for now, I am not just sniping. I have a full artificial general intelligence research program described in my book: https://mitpress.mit.edu/books/algorithms-are-not-enough. I don’t claim to know the answers, but I do claim to know at least some of the questions.

That is why recently people have moved towards very large neural nets, where representations arise implicitly. Such representations are shallow, and do not separate the data from the concepts, but are more flexible.

I think learning how to build representations reliably is very hard. A solution where we guide neural nets towards building them based on data is likely going to work better than us doing that.

Yes, current approaches are very limited. But they are good at solving concrete problems, and the range of problems it can solve is getting larger.

Any time people tried to do clever things they failed. Only pragmatic focus on incremental improvements worked.

We will move beyond current algorithms, but incrementally, as we uncover more patterns in the world and as we see more systematic problems with current methods.