tcdent

Yeah, it's such a subjective feature. I don't really worry about the five minute timeout, but sometimes if I'm conscious about the fact that I have an 800k token context window that I haven't interacted with for a day, I will switch to Opus for compaction and then back to Fable for interaction.

In general, though, some kind of API ping on a timeout does not support my personal workflow, which involves dozens of active or stagnant agent sessions that stay open for weeks.

The ambiguity of some of these supports my read btw.

There might not be data to support it since it touches on legally sensitive subjects, but the deaths are almost entirely attributed to drug use.

Deaths due to overdose at music festivals in general are not uncommon and are kept quiet (somehow).

My read is that it's generally the teams that have been assigned to build a certain product that end up choosing the architecture that it runs on. So when we see TypeScript involved in TUI and CLI applications, most of it is just a repurposing of skill sets from that domain into the terminal. In an organization like Cloudflare, I expect that the developers who don't specialize in TypeScript are working on far more important problems.

The real reason why plan mode is dead is because you can just conversationally instruct the agent to not make changes to the repository or to make changes to selected documents only, and it will listen. There was a time when we needed to enforce this via selected tool use, but we have surpassed that.

Yeah, it's way better when you do design documentation, or even ticketing, to instruct it not to include any implementation specifics. You're not doing the deep dive on the zero shot that writes the ticket or the document and so it is much less informed than the agent doing the work will be.

The DoD has been using AI for engagement since the 80s. Maven, the system used in your example has been active since 2017.

None of those systems pull the trigger. They assemble data for human review.

A poor track record compared to who? The Romans?

War includes death by design. The trajectory is trending toward less unjustified death.

> [Anthropic] would not permit its technology to power fully autonomous weapons without human oversight over targeting and firing decisions. Defense Secretary Pete Hegseth argued that the Pentagon should have unrestricted access to AI systems for “any lawful purpose” and that a private contractor should not be permitted to dictate how the military can use its technology.

It's about them maintaining an interpretation of law at all.

Simpletons read that as: "The DOW wants to kill people with AI."

https://www.pearlcohen.com/anthropic-sues-department-of-defe...

On faltering: the US Military already has rules around the use of automated systems for this type decision making. This is not a new concept to them by any means, and currently it always involves a human decision.

As I understand it cane down to one point: Anthropic wanted the ability to review actions that would have an effect on human life.

Which sounds sensible on the surface, until you realize that all military organizations have extensive experience with that exact question, and have spent centuries refining their practices. Deferring that to a private company is actually a major regression and a concerning role to pass on to the public regardless.

FWIW I am not pro military, but it's a prominent part of human culture and not going away anytime soon. I do understand the nature of professions, however, and only one of these groups has made the interpretation of this question their specialization.

I'm generalizing the political groups involved, but I don't understand how anyone who is conscious and/or cautious of AI safety, perhaps to the point of regulation, would then support the usage of AI-assisted surveillance.

Multiple things can be true at the same time.

You would think a website dedicated to talking about software engineering would agree with this sentiment wholeheartedly.

A dense model (up to the amount of memory available) actually does make the most sense on unified memory architectures. But when you hit the limit of what you can hold in memory, you reach the limitation of the platform.

Whereas a hybrid architecture with distinct DRAM and VRAM with sparse MoE, you can leverage two different bit rates depending on the actual need for constant access to common layers versus sparse access to infrequent layers and arbitrage the difference in cost for each of those in distinct classes of hardware.

A query parameter on a GET request is actually data written to memory. So this idea that any HTTP verb somehow provides context or enforcement of read versus write totally misses the point.

So many different theories on why this is.

First, I think comprehensibility is a major part of why certain products grab the interest of the mainstream portions of the market. The 75% of posts in your feed are not from people who evaluate products based on underlying technology. They typically value signal from social reinforcement higher than anything else. This is the same reason why we see people mentioning products instead of technologies, i.e. PlanetScale versus Postgres & Tailscale versus WireGuard. The consumers understand the value proposition, but would have never discovered it without relatable messaging. This isn't a new phenomenon in computer software either; jQuery is probably one of the first examples that I can remember with this sort of texture.

The other side is a perception of expertise in a specialty. Software development, especially in AI, has become an incredibly desirable profession, and there are more people than ever racing to be included in it. In my own professional experience I find an excessive amount of entry level talent leveraging the same comprehension of product, but not comprehension of technology to get their foot in the door. A vast majority of the "thought leaders" occupying our feeds are not as well practiced as they claim to be, they're just trying to get a job or raise funding.

And finally, AI has brought out a certain amount of desperation in practitioners, for lack of a better term, materializing as an anecdotal, but certainly observable need to remain on the very tip of the news cycle in order to feel well informed. And so, using the dynamics above and many other human social dynamics, we find certain concepts spreading across cohorts that would not normally have a need or a want for these particular techniques, or products, or solutions, but because they feel pressured to remain relevant.

It's essentially taking output schemas as we've been using them and applying them to specific classification tasks. So not using them to generate structured content which incorporates generated text, but using them to generate structured content which includes classification and/or rankings of the requests made.

So in a lot of cases when we've used LLMs as a classification hack, we've burned a ton of tokens in reasoning and output that we didn't really need to use to interpret the final result. (And I'll just say that we may not have needed all of the output tokens, but that incorporating assessment along with scoring seems to provide more accurate results.)

This goes beyond just asking an LLM to assign an arbitrary number to a particular concept, which in most cases distributes less-than-correct statistically, although that didn't stop us from considering LLM as a judge to be a viable strategy.

So this basically gives us a different class of model to use when classification or decision making is the only need. It doesn't replace any of the narrative if you still need that. Coupled with the higher speed and lower cost, that's why everyone's excited about it.

I don't think I quite valued the Open Source PHP ecosystem at the time as much as I should have. I have a project that still grabs a ton of installs for some reason (3+ million to date and apparently growing) which is way beyond anything I would have expected.

https://packagist.org/packages/tcdent/php-restclient/stats

I reached out to their chatbot after my battery life tanked a year-or-so later and it sent me a replacement no questions asked.

Is this "pacing the frontier"?

This is the part I find being left surprisingly hand-wavey.

If you extrapolate it out, you see that it has a high likelihood of inciting physical force (read: military action) as a means of enforcement, so perhaps that's why nobody promoting ideals has been direct about it.

Read other writing by Anthropic about potential future financial implications of AI. [1]

IPO is a vehicle for future wealth distribution. It leverages existing financial frameworks in order to span the gap from before-to-after without having to convince the system that future is inevitable.

[1] https://www-cdn.anthropic.com/files/4zrzovbb/website/9ea607a...

OP has practically observed the effect of the Ego via software, but hasn't quite arrived at the ability to quantify it.

Read back through and apply every example given through that lens.

It is possible that we're in an industry that inordinately expresses this part of human nature, but I'm pretty sure it shows up everywhere though different anecdota. Apply a reductionist Zen Buddhist view to your professional creativity and all of this goes away.

Bob wants to refactor a subsystem because it will make him a hero, and if he positions it correctly to management, the technical merit and actual realized level of success will be irrelevant. Alice chooses to surface some obvious concerns and then sit back and watch the show. Jane chooses to throw her hands up in stand-up and try to emotionally convince everyone the sky will fall. Be like Alice and preserve your sanity.

But here's the thing, across operating systems the products should not be identical.

When you get to the point that your have a significant enough number of users across multiple platforms, generalizing everything into a shared UX doesn't make sense; giving those classes of users the best experience requires embracing their platform.

A simple example: Android has a system-wide convention for a `back` button. iOS has no such standard. Users on each platform have different expectations for how to navigate an app fundamentally, and holding tightly onto the concept of identical gives both camps a compromised experience.

Came here to say this as well. It's the obvious all-encompassing free office document processing suite and LLMs know how to utilize it well.

Codex bundles it with the desktop app as well for the same purpose.

I had an LLM reverse engineer the API protocol of LADBS and extract the content I wanted. Wether that scales from one property to many (rate limits for one) I don't know, but I didn't encounter anything in the thousands I extracted.

I did a project recently where I extracted every Los Angeles building permit record for a property and was able to look at the evolution of its infrastructure and tenancy over time by embedding the permit documents into an image-based embedding model. Could be an interesting way to incorporate more granularity of the history of properties over time, although it dramatically increases the scope of data processed.

V8 as a runtime goes far deeper than just webpages.