evmaki

IME waiting for an agent to work through a problem is a detriment to attention span; your mind drifts to other things while you wait. Maybe you can steer several agents in a round robin instead, but then there's a cognitive tax from context switching. Faster models mean fewer gaps in focus.

Nonetheless, this is very cool work! If I can offer a small suggestion to the team at Cactus, it would be to evaluate your releases on some usability criteria (including false positives). Any serious integrator or adopter of these models would want to have that information available.

Also appreciate the attention to detail with licensing in the training set. This is an important sticking point – both commercially and ethically – for any product that integrates this type of model.

AI is just the next wave, and the impact is more tangible than ever – it literally takes your job, and it's being pushed on you by enormously wealthy people who don't understand you, your life, and what's important to you. The sad thing is, AI can be beneficial to people if wielded in the right way, but we are in a polarized environment where productive conversations no longer feel possible: you're either an AI bro or a luddite. I think anyone (myself included) who has spent time developing B2C products that incorporate AI quickly discovers just how touchy of a subject this is, and it's due imo to the sins of the accelerationist crowd that never wastes time understanding the needs and perspectives of normal people.

https://en.wikipedia.org/wiki/List_of_college_radio_stations...

^ One of those might be in your area :)

While these tools stand to enable the democratization of productive capability in software engineering and other tasks (creating a renaissance for solopreneurs, let's say), what seems more likely to actually happen is that entrenched capital will become the only player with real access to this "knowledge as a utility" (was it Altman who called it that?).

We already see this playing out in two fronts: 1) the gradual reduction of services and 2) the DRAM market, where local-first tools (i.e., potential disruptors of the emerging "knowledge monopoly" created by the big AI firms) are being stifled by supply shortages. How many promising small-to-medium-sized competitors are being snuffed out of existence (or never starting) due to the insanity of the DRAM/storage/CPU (soon) markets?

The currently-subsidized access that we have to the big Opus-like models will, in parallel, be gradually be taken away until only the big players can afford it. And in the end what we will have is hyper-productive skeleton crews at a few consolidated firms performing (or selling expensive access to) basically all of the knowledge labor for society, with very little potential for disruption due to the hardware and "knowledge" scarcity engineered (in part, maybe) by this monopoly.

Not necessarily a closely held belief – just a hunch – which is why I want to see what parts of the picture I might be missing.

I've been wondering if it is even possible for a publicly-traded company to deliver a voice assistant product without these incentives involved. I have to imagine the UX of these devices would be much different if they were built by a private company without the same market pressures. It would need to be self-contained and local, so that the infrastructure burden (e.g., data and AI in the cloud) wouldn't create a need for subscription service or data collection revenue to cover the cost.

I'm fairly convinced a big part of clearing the barrier to entry of these elite institutions is having a deep understanding of exactly the things you need to do to succeed given the structure of the system and the nature of the competition. Students at "non-elite" institutions are more likely to come from backgrounds where even if you DO have a disability, maybe nobody ever tells you that you can go to the doctor for it, or that something like "accommodations" exist to help you.

We are probably undervaluing the human part of the feedback loop in this discussion. Claude is able to solve the problem given the appropriate human feedback — many then jump to the conclusion that well, if Claude is capable of doing it under some circumstances, we just need to figure out how to remove the human part so that Claude can eventually figure it out itself.

Humans are still serving a very crucial role in disambiguation, and in centering the most salient information. We do this based on our situational context, which comes from hands-on knowledge of the problem space. I'm hesitant to assume that because Claude CAN bootstrap skills (which is damn impressive!), it would somehow eventually do so entirely on its own, devoid of any situational context beyond a natural language spec.

https://en.wikipedia.org/wiki/Incarceration_in_the_United_St...

https://en.wikipedia.org/wiki/Incarceration_in_the_United_St...

More speculatively, I think the prison system has also taken over the role of the mental health institutions that were wound down under Reagan. Over half of the incarcerated population has a mental health condition, and likely are not receiving adequate mental health care while incarcerated.

Best of luck with your Kickstarter!

On the other hand, I wonder if this doesn't just further alienate children from their parents. Kids are already given access to unlimited supernormal stimuli via iPads so that parents don't have to parent. This just seems like more of that: now parents don't even need to have basic conversations with their kids because the AI can do it.

Anecdotally, some of the most formative interactions I had as a child started by asking my parents questions. These were things that not only shaped me as a person, but deepened my relationship with my parents. These interactions are important, and I wonder if Aris doesn't just abstract it away into another "service" that further deepens social decay. I would not be the person I am today if I hadn't had the chance to ask my dad as an angsty pre-teen what the point of life is, and for him to tell me it is to learn and create so that we can make a better world for humanity. I guarantee a smoothed-over LLM would not have offered something so personally impactful.

My two cents is that you should ponder that deeper point a little bit, and think about how it informs the way you market your idea, and scope the service it provides.

Is that inherent to the technology, or is that just inherent to the way we've chosen to organize society? Really, any technological paradigm shift going back to the industrial revolution has mainly served to enrich a small few people and families, but that's not some immutable property of technology. Like you say, it's a tool. We've chosen (or allowed) it to be wielded toward one end rather than another. I can smash my neighbor's head in with a hammer, or I can build a home with it.

At one point in the United States, there was political will to update our social structures so that the fruits of technological and economic progress did not go disproportionately to one class of society (think early 20th century trust busting, or the New Deal coming out of the Great Depression). I'm afraid we find ourselves with a similar set of problems, yet the political will to create some other reality beyond further wealth concentration seems to be limited.

I've been working on this problem in an academic setting for the past year or so [1]. We built a very similar system in a lab at UT Austin and did a user study (demo here https://youtu.be/ZX_sc_EloKU). We brought a bunch of different people in and had them interact with the LLM home assistant without any constraints on their command structure. We wanted to see how these systems might choke in a more general setting when deployed to a broader base of users (beyond the hobbyist/hacker community currently playing with them).

Big takeaways there: we need a way to do long-term user and context personalization. This is both a matter of knowing an individual's preferences better, but also having a system that can reason with better sensitivity to the limitations of different devices. To give an example, the system might turn on a cleaning robot if you say "the dog made a mess in the living room" -- impressive, but in practice this will hurt more than it helps because the robot can't actually clean up that type of mess.

[1] https://arxiv.org/abs/2305.09802

To your point about turning random commands into API calls: if you give it the raw JSON from a Philips Hue bridge and ask it to manipulate it in response to commands, it can even do oddly specific things like triggering Hue-specific lighting effects [0] without any description in the plugin yaml. I'm assuming some part of the corpus contains info about the Hue API.

[0] https://evanking.io/posts/homegpt/

I have been working in this area (LLMs for ubiquitous computing, more generally) for my PhD dissertation and have discovered some interesting quality issues when you dig deeper [0]. If you only have lights in your house, for instance, the GPT models will always use them in response to just about any command you give, then post-rationalize the answer. If I say "it's too chilly in here" in a house with only lights, it will turn them on as a way of "warming things up". Kind of like a smart home form of hallucination. I think these sorts of quality issues will be the big hurdle to product integration.

[0] https://arxiv.org/abs/2305.09802