This Anthropic post confirms what I’ve been seeing with MCP: loading all the definitions upfront burns tokens fast. It also buried the lede that MCP creates context rot. There's an easy fix but it requires us to do actual engineering rather than spray and pray: progressive disclosure through a good router. Your agent gets exactly what it needs, when it needs it, without paying the context or token cost for unused tools.

the best movie on context engineering

This is sad to hear. Learn new tools people! This is by far the most exciting time in software land in the last 20 years.

I'm quite careful about "attention hygiene" during my working hours, but one thing this taught me is that "poor hygiene" in off-hours (even the previous day!) can really impact the quality of my awareness.

For a few years, I've had weekly 2-3hr readings for a course. It's a startling barometer: if I'd spent a while on X/YouTube/etc in the prior 48 hours, I'll want to distract myself every few minutes. But if I've been scrolling-free, it'll be easy to settle in for an unbroken hour!

Steve Newman gives a great overview of the gigantic factual disagreements underlying a lot of disagreement about AI policy

France is building 220km of high speed rail for the same cost as San Francisco building 3km. Transit costs are one of the biggest barriers to America building world-class sustainable infrastructure, and we need our politicians to enact policies to lower construction costs.

This from @elidourado is still the best articulation of the specifics for an abundance agenda in Washington, DC

I'm not an effective altruist, but I find the recent genre of pieces like the below a bit strange. Essentially all of the major AI lab leaders agree that AI potentially poses enormous risks[1], as does a majority of the US public[2]. It's not a crank view. EA was simply among the very earliest organized groups to perceive and act on this risk. In addition, EA's contrarian concerns have elsewhere been shown to be reasonable: long before COVID, EAs stood out for their worry about major pandemics. Assessments that deride EA as a cult while failing to acknowledge these counterintuitive successes strike me as unreasonably uncharitable. If asked "which 2018 community now looks most prescient based on how the intervening 5 years unfolded?", I think it's hard to come up with better nominations than EA. politico.com/news/2023/12/30… [1] nytimes.com/2023/05/30/techn… [2] rethinkpriorities.org/public…: a majority rate the proposition that AI will cause human extinction within 50 years as at least "slightly likely".

some excellent takes from @labenz on the latest episode of the @CogRev_Podcast that I just wanted to draw attention to is the development of AGI really inevitable, and the singular way to advance human civilisation, or are these deeply ideological claims?

AFAICT everyone involved in this is both ridiculously smart and a good person and all deserve our love, respect, and sympathy for what is an incredibly stressful time.

This whole thing still happened for a reason, and OpenAI needs to do a much better job explaining it. If internal concerns about upcoming capabilities — and how they were handled — really came to this boiling point, people need to know.

This all seems like a clever plot to finally get me to pay for The Information.

The whole OpenAI conversation is so infuriatingly stupid right now. The OpenAI board needs to say was Sam was fired for right now. Period. Full stop. If it was Ilya not liking being demoted, GTFO. If it was them not liking Sam raising money for Nvidia competitors and AI phones, reasonable people might disagree but even those who view it as reasonable will have to admit it was massively bungled. If it was because of GENUINE fear of the consequences of specific actions that betray the nonprofits mission, they need to be specific about what those actions were and where the fear emanates from. And I'm sorry, they made this bed so hiding behind anything along the lines of "it would be too bad for the publicly to know" won't cut it. The harm that the board is causing by NOT getting specific -- to the company AND to any sort of safety mission -- is hard to overstate.

Since I am now somewhat of a journalist in such matters, I should note that if anyone wants to provide any info (on or off record) on recent events at OpenAI, DMs are open.

It turns out that it’s hard to control the actions of a highly intelligent, hard-working, and charismatic agent.

lots of predictable takes and hypotheses that magically align perfectly with posters' preconceptions and narratives about safety or accel or open source or whatever they care about. in the absence of any information maybe just wait

Manhattan from 100 years ago looks like SF's Sunset and Richmond today. Imagine of Manhattanites back then had said "Sorry, we're built out!" That's what SF has been doing for the past few decades.

The reward for reading 20k words of an EO that is almost entirely calls for government reports is to read the diatribes about how it is government overreach and full of premature regulations and assumptions, when you spent hours confirming that nope, it's all report requests.

Differential accellerationism: 1. Go slow when the downside risk is millions of deaths (or worse). 2. Go fast when it isn't. The world can get so much better if we really go for it on both (1) and (2).