clbrmbr

I think so, if you can inject synthetic errors with the same distribution as the errors you care about.

It is surprisingly hard to do in a single prompt. I’ve had good luck though with being very explicit about asking for review, then to think deeply about what actually matters, and then reduce to a very short extremely concise reply. —- the model NEEDS to output those pages of text as part of its thinking process. It’s used to doing it in the assistant reply but can be cajoled into doing it in the thinking tokens.

I’ve found that doing the full requirements capture, planning, writing, reviewing, gardening loop with frontier models has worked quite well since last October, and phenomenally since Fable 5.

The key I’ve found is human peer review. The reviewer jumps on a live call with the developer, pulls up the PR with transcription on, and asks questions. At the end of the call, the transcript passes back into the coding agent and the PR is polished up, becoming more self-documenting, and the humans are left with some degree of common understanding of what’s going on.

I’ve been operating my team of ~15 this way for 9mo to great effect… there is simply no going back to the stone ages.

Mais par contre, Canada also has some tremendous senatorial committees. They asked real questions of Hinton.

Matches my experience. Basement home offices are the worst offenders.

Do we have recent commentary on open-weight from the Anthropic founding team?

With $onnet he would have gotten pwned. Or at least I’d love to see a comparison against other models.

It is customary that one may publish one’s own personal correspondence unless the other party has requested confidentiality.

Maybe this open invitation to the world pushes the boundaries of that definition, but I don’t see where an expectation of privacy comes in here.

Hmm I imagine using a server to connect to signal/whatsapp or even email, then using a local model to classify and filter and trim messages and forwarding to SMS, and viceversa. I guess the trouble is I’d need many source numbers :thinking:.

Fascinating. Can you explain why southern London is DC while northern London is AC?

Can you share an example? I've been happily using Fable this afternoon and it just seems like the usual upgrade so far with no interruption to my (fairly standard) SWENG problems.

Very nice. Can I fork for my private use? Was thinking of building similar.

Philanthropically-minded people will move the winners to a donor advised fund which gives FMV write off without ever paying capital gains.

With index funds you never have the strong winners to do this with, and so giving is far less tax-efficient.

But do they historically beat the S&P 500?

Which suburb, haha. I’m in Englewood and have similar experience of very few tech folks around.

Was a good watch, tho would have liked to be there in person. Props to Brenden & his Cosmos team for really setting the bar.

One capability that I see is missing from opus is this ability to understand an entire system. My hope is that a mythos class model will be able to comprehend even something as complicated as an IOT system with a hardware and firmware layer multiple API’s backend and different kinds of API and web clients.

The main limitation we’ve had to agentic coding is an understanding of this system that spans processes running on different machines and architectures.

This. I added that instruction the first and last time I was gaslit by an underpowered subagent.

> You asked a simple question. They lobbed a document.

I’ve become so cynical that I cannot read this phrasing without thinking it came from… an LLM.

This cynicism also negatively impacts human communication. This constant doubt of whether there was a thinker on the other side of the text.

Come on, Anthropic ARE the good guys if there are any. Certainly the incentives of trillions will do what money does, but they have assembled an incredibly altruistic and philosophically-minded crew. I’m rooting for them and trying real hard not to get cynical.

Karpathy embedded within an organization is way more impressive than him out on his own with hot takes and little projects. I hope he does great things for Anthropic.

Anyone here using OpenBSD? If so, for what purpose?

I’ve always wanted to use NetBSD for an application for an embedded system / IoT device but never had the pleasure (yet!).

Thanks was a good watch. Sad though the example of the AI app to “help farmers” that is making things up. I would expect a generational cassava farmer to have a much better sense of how to treat the plants than an image model.

This mirrors my experience at Stevens. The professor would not babysit us during exams and that really did inspire pride. Also the exams were often brutally hard which inspired despair.

High-end Chromebook done right could be a very good thing for computer security.

I thought similarly, until I actually tried using AI to shop for clothes, now I’m a total convert. It’s like the best possible men’s fashion concierge…

Do you know if this exploit works on Docker containers? And if so, I assume it just allows escalation WITHIN the container? So this attack is scary for Linux desktops and servers, but a fully containerized system like common on CI/CD should be good. Right?

It’s interesting to compare how the agentic search performs, with these targeted reads and lots of tool calls in the stream, versus the older but still valid paradigm of using a high-reasoning model like GPT-X-pro and feeding in all the relevant files at once with no tools.

I have found that the “pro” approach is much more holistic and able to tackle rather “creative” problems that require very careful design and the overall artifact is tight and self-consistent. — Claude Code by comparison is incredible in exploration and targeted implementation but indeed is not great at seeing the forest.

So what do we do? Pin our dependencies (to hashes when possible), and only update when there are CVEs?

But problem is this could lead to abuse of the CVE system to try to force rapid adoption of attacked packages. What prevents this?

Wouldn’t you prefer to pin to SHA hashes? Or does your package manager cloud-side ensure immutability of releases?