searealist

OpenAI's view of multi-agent is something like: you get the same work faster, but it does not scale linearly.

Only you can decide if that speed is worth the cost.

MLK's entire organization was largely created and run by Stanley Levison, who was a big player in the CPUSA. He was a major financier of communist efforts in America before switching his focus to the civil rights movement.

But you won't find Levison's name on the SCLC Wikipedia page even though his personal wiki page mentions he was "instrumental in founding" it.

> Does anyone understand why they are calling this a distillation attack?

I once had my credit card stolen and it was used to purchase gift cards.

Interestingly, no one referred to it as a “gift card attack”.

Between "distillation attack" and "under attack by agent swarms", it's probably to absolve themselves of responsibility by making it seem like an unavoidable arms race.

I wouldn't expect a post-mortem anytime soon.

> security breach

> related to a large scale distillation attack

I'm not sure how this adds up. I could understand if this impacted their uptime, but not security.

Are they actually saying a bunch of credentials were stolen and are being used for distillation?

Yarvin is about as anti-rationalist as you can get.

As someone who has played BAR, it's hard to read this as not being maximally sarcastic.

The cancer thing is just a contrived analogy. His real concern is having himself and his mathematician friends replaced by AI. Appealing to someone's health is an easier sell and then it can normalize the relationship of medical priests, and math priests, etc.

Can we skip to the moat?

It's more like this:

Right now when a model wants to call 1 or more tools, there is a fixed json schema to describe the tool calls.

Codemode is like: Why don't we just let the model write a program to call the tools and compose them however it wants? The key is that the programming language exposed to do this (usually javascript) will have APIs available to do some things internal to the harness (like call tools/mcp).

That's one kind of MCP. Another is a local stdio server.

Also there are things like subagents, etc (which may be considered tools).

Can do you make a tool or mcp call from bash?

In Canada?

There are good reasons they have all converged on this model:

- The models can generate html/js as output instead of markdown. They all support something like "artifacts" that can displayed in the transcript or a tabbed panel.

- Many of them are able to have the model generate plugins to customize the harness. Typescript/V8 is really good for this.

- They all allow you to open web pages in a tabbed panel. Sure, you can do this with native be embedding a web browser, but with the previous points, you have to wonder if you should just use Electron / Tauri.

I dunno. Windows 98 is very usable. What does Gnome 3 have that someone who never used Windows 98 would miss?

What other desktop environment doesn't have a dock or taskbar?

Age has nothing to do with it. Gnome 3 is just bad.

You can configure KDE Plasma to be nearly identical to Gnome 2 / Mate.

It's great.

Unlike modern Gnome, KDE Plasma is _really_ good. In fact, you can configure it to almost perfectly mirror Gnome 2.

That being said, this is really cool, and brings back a lot of memories.

On the one hand, you have a meta-analysis estimating 7–28% of the U.S. homicide decline, depending on the model, after correcting for publication bias. That’s 4–15 percentage points of a 54% decline.[1]

On the other hand, you have the Dunedin study, where lead exposure was uncorrelated with measured socioeconomic status, which found weak links with offending overall and no significant link to violent crime.[2]

In any case, people talk about this like lead was the dominant factor in the violent crime decline, when these results don’t establish that. Even the meta-analysis leaves most of the homicide decline unexplained by lead.

[1] https://eprints.gla.ac.uk/332830/1/332830.pdf

[2] https://jamanetwork.com/journals/jamapediatrics/fullarticle/...

I thought we were pretending it was lead?

How does this relate to 4500 vs 5090? I'm just pointing out that 5090 likely has twice the performance of the 4500 and likely maintains that at 2x watts if you want.

A 5090 has 2x tensor cores and 2x bandwidth and can be run at 400W (2x watts).

... or with llama.cpp with MTP.

It depends who is president.

2x is more typical, and only for dense models, not the sparse models all frontier labs use.

Also, even the best speakers in the world can benefit from room correction, because every room behaves differently. The difference is HUGE in the lower frequencies.

Don't forget about electricity costs, esp in California. Even with free hardware you may be better off using APIs.

Aaron had a history of depression and suicidal ideation that long predates his prosecution.

Also: https://www.unqualified-reservations.org/2013/01/noam-chomsk...

Swartz was federally charged with wire fraud and violations of the Computer Fraud and Abuse Act based on allegedly unauthorized access, not simply prosecuted for copyright infringement or “downloading articles.” Also, he was offered a plea deal of 6 months and his own attorneys did not expect him to serve any time even if rejecting the plea deal and convicted.

Meta is accused of civil copyright infringement.

Very different things.

> In my experience MTP's speed increase doesn't seem to justify the apparent loss of success at the edge, it would have to be at least 4x faster to meaningfully churn through the first 3 failures in the time it would have taken to do it once without

Speaking in terms of wall clock, the expensive part of decode is fetching the weights from memory. Predicting and validating a bunch of tokens using the already fetched weights is insignificant in comparison. Even if you have a poor acceptance rate for predictions, you won't really see a slowdown vs not using MTP.

Just enable MTP on llama.cpp and you will get the same decode speeds.