manbash

Oh, definitely. Also, they end up getting stuck in a loop, adding and removing the same code endlessly.

And then someone comes and "improves" their agent with additional "do not repeat yourself" prompts scattered all over the place, to no avail.

"Asinine" describes my experience perfectly.

Aw shucks I was almost pulled for another round of Star Fox 64. Unfortunately it's not there.

Ah, those days, where you would slice your designs and export them to tables.

How else can you take responsibility if you don't make it public? You can't have integrity if you hide away your faults.

Under "Programmatic Tool Calling"

> The challenge

> Traditional tool calling creates two fundamental problems as workflows become more complex:

> Context pollution from intermediate results: When Claude analyzes a 10MB log file for error patterns, the entire file enters its context window, even though Claude only needs a summary of error frequencies. When fetching customer data across multiple tables, every record accumulates in context regardless of relevance. These intermediate results consume massive token budgets and can push important information out of the context window entirely.

> Inference overhead and manual synthesis: Each tool call requires a full model inference pass. After receiving results, Claude must "eyeball" the data to extract relevant information, reason about how pieces fit together, and decide what to do next—all through natural language processing. A five tool workflow means five inference passes plus Claude parsing each result, comparing values, and synthesizing conclusions. This is both slow and error-prone.

Basically, instead of Claude trying to, e.g., process data by using inference from its own context, it would offload to some program it specifically writes. Up until today we've seen Claude running user-written programs. This new paradigm allows it the freedom to create a program it finds suitable in order to perform the task, and then run it (within confines of a sandbox) and retrieve the result it needs.

> I speculate that within a few months, the communities will have settled on a single dominant one.

The solutions on the roadmap are not centralized as GitHub. There is a real initiative to promote federation so we would not need to rely on one entity.

Nice work. I have gone through the fairly straightforward paper.

May I ask what you've used to confirm the cache hit/miss rate? Thanks!

I agree with you, but also I want to ask if I do understand this correctly: there was a paradigm in which we were aiming for Small Language Models to perform specific types of tasks, orchestrated by the LLM. That is what I perceived the MCP architecture came to standardize.

But here, it seems more like a diamond shape of information flow: the LLM processes the big task, then prompts are customized (not via LLM) with reference to the Skills, and then the customized prompt is fed yet again to the LLM.

Is that the case?

Right in the first paragraph.

> Needless to say, I just don’t get git.

What is there not to _get_, honestly? And why is jj so easier to get?

The author seems to focus on how great it is to make changes to your commit history locally, and that you shouldn't worry because it's not pushed yet.

The thing is, I don't want automatic. Automatic sucks. The point of version control is that I am able to curate my changes. The guards and rails of git is what makes me feel safe.

I am still failing to see why JJ is superior to git, or whatever.

Don't you have a 2FA Recovery Code?

> They don’t prevent the attacker from pivoting a memory safety bug to remote execution.

I'm confused. Isn't this potentially preventing some classes of memory-safety bugs?

> Not all free software is open sourc

Depends on which "free software" definition you're referring to.

The FSF definition of "free software" requires it to be open source.

Maybe a "view", however YT defines it, is a poor metric? It doesn't show engagement and it's time for YTers to have better means for analytics.

I have an Android TV device, and YT has been so horrible with its constant ads popping up, that I have to put it on MUTE to prevent any further brainrot.

I wonder when they're going to blame me muting my TV and harm their viewership. Or maybe they will just prevent me from being able to mute it.

Do they charge for actual bandwidth as well? Seems like it. From tarsnap.com:

> Tarsnap uses a prepaid model based on actual usage: Storage: 250 picodollars / byte-month of encoded data ($0.25 / GB-month) Bandwidth: 250 picodollars / byte of encoded data ($0.25 / GB)

Thank you. I assumed it would be quicker to find the link to the case PDF here, but your summary is appreciated!

Indeed, it is not only payout, but the destruction of the datasets. Although the article does quote:

> “Anthropic says it did not even use these pirated works,” he said. “If some other generative A.I. company took data from pirated source and used it to train on and commercialized it, the potential liability is enormous. It will shake the industry — no doubt in my mind.”

Even if true, I wonder how many cases we will see in the near future.

Thanks for this. I was struggling to put it in words even if maybe this has been a known distinguishing factor for others.

I tend to agree, but in a more general and broad sense. That is, Rate-Limiting is always a deployment specific feature. If it's for limiting user requests, then it should be a component of the ingress/API Gateway and be as robust.

> They just aren't distributing the source for free, it seems, but you are free to redistribute it however you'd like.

Yes, that was my point. I will know the license when I see it in the distributed code :)

This looks great. Just a note though that this is "free as in beer". If you want the source code you'd need to login and pay, it seems.

https://studio.blender.org/projects/dogwalk/gallery/?asset=8...

> 2. Technical design based on requirements > Kiro then generates a design document by analyzing your codebase and approved spec requirements. It creates data flow diagrams, TypeScript interfaces, database schemas, and API endpoints—like the Review interfaces for our review system. This eliminates the lengthy back-and-forth on requirements clarity that typically slows development.

This is nice for documentation but really having a design document after-the-fact doesn't really help much. Designing is a decision-making process before the code is written.

I am reminded of a theory years ago that fingers wrinkling in water might have an evolutionary function, specifically to improve our grip when wet (or in water).

The wrinkly formation lets water drain better (like treads of a tire).

Are you able to look around while keeping your "unified vision"?

To me, all the differences appeared to be flashing (probably my brain alternates between the pair of images it attempts to "lock in", or something to that effect).

I wouldn't be concerned about trusting the results of ChatGPT if it also were providing links to the sources it had cited or used as a reference in its answers.

Unfortunately, it doesn't, and so I can't verify them. Not sure if it's an actual limitation of current LLM or rather they're intentionally filtering out the sources.

Is Stepmania still the go-to DDR clone? This is very nostalgic. :)

VPN is used to bypass regional restrictions.

The VPN provider will share information if an active investigation is underway.

I sometimes indulge in doing the KFC chicken rice cooker hack recipe. Look it up if you haven't yet.

I do too wonder.

Given the disruption of LLM everywhere, and in Google specifically, I wouldn't be surprised if the headcount was reduced in these teams as well.

A side note: I am suspicious about the channel in the posted link. It seems to simply cut clips from the original video, without information and credit.

Anyway, here is the original source: https://youtu.be/cPvRIWXNgaM

State diagram and sequence diagram complement each other, rather than competing.

I am surprised by the lack of eSIM option, which unfortunately is a dealbreaker for me. :(