
I like making things with others. LLM shepherd. Godโs first commandment to Abraham was to smash all the idols. https://xitter.cf/t.co/pCDpWgOOSz (Header photo by @cartoonconnie)
Joined March 2008
- Tweets15.5K
- Following4.9K
- Followers2.6K
- Likes16.4K
Pinned Tweet
For almost the whole last year, Iโve been writing and polishing an article that lays out a frameworks for how to think about (and fix) social media. And now itโs finally here, thanks to @techreview. Hereโs a thread with some key points ๐งต technologyreview.com/2021/12โฆ
This is the way
ืืช ืืืดืจ ืจืขืดื ืืดืจ ืื ืกืืจ ืขืืืก ืื ื ืืืืจ ืืคืืืฉืืช ืืฉืืืืช ืืจืืืืช ืืื ืกืช ืืืืืฆื ืื. ืื ืกืืจ ืืื ืื ืืื(!) ืขืจืื-ืืฉืจืืื ืืืืฅ, ืฉืืืืืฉ ืืช ืขืืืชื, ืืื ืฉืืืืจ ืืืืื ืช ืืฉืจืื ืืืืื ื ืืืืืืช ืืืืืงืจืืืช ืืจืื ืืืืืช ืืขืฆืืืืช, ืืืืื ืื ืืืืืก ืืื (ืืื ืื ืืฉืืจืืช ืฆืืื ืื ืืืืื-ืืืจืื). ืจืืืชื ืืฉืืขืชื ืืืชื ืืืืื ืณืฉืืืจ ืืืืืชืณ, ืืืฉืจ ืขืืจ ืืขืจืื ืืืขืืจืืืช ืืดืฉืคื ืืืืด ืืืงืืืืช ืืื ืืืื ืืกืืื ืฉื ืื ืืืืจ ืดืืฆืืชื ืืฉ ืฉื ืฉื ืื, ืืฉืคืื ืืืง ืืืืืจืืด.
ืืช ื ืืฆื (ืืืืืืก) ืืืื ืกืืืืืืฅืณ, ืื ื ืืืืจ ืืืื, ืืื ืฉืฉืืจืช ืืคืงืืื ืืกืืืจืช ืืฆื ืื ืื.
ืืช ืืืชืืจ ืื ืืืืจ ืื ื ืืืืจ ืืืืืงืื ืฉื ืืฉืืดื, ืฉืืืจืืจ ืฉื ืื, ืืกืชื ืืืืืืืช ืืืืจืช ืืืจืื. ืขื ืืขืฉืื ืืืจืฉืข ื-8 ืืชืื ืืืฉืื, ืืื ืืื ืชืืืื ืืืจืืื ืืจืืจ (ืืืืื).
ืืช ืืฆืืื ืกืืืืจืืฅืณ ืืืจืชื ืืื ืฉื ืขืฆืจ ืืคื ื ืืืชื ืชืงืืช, ืืืื ืืฉื ืืืืื ื ืืคืืฆืฅ 700 ืืืืจ ืืืง ืขื ืืืืฉ 1 - ืคืืืืข ืฉืืืืจ ืืื ืืขืฆืืจ ืืช ืืืฆืืข ืืืชื ืชืงืืช. ืขื ืืขืืืื ืืฉืจ ืืืฉืจื ืืืืืืื, ืืขืืืื ืืืืืื ืืืจืืจ ืืืืืื ืืืืืื, ืฉืืืจืื ืืืงืขืช ืืืจืื, ืืื ืฆืืจื ืืืจืืื.
ืื ืกืืจ ืขืืืก ืืื ื ืื ืืช ืืฉื, ืื ืื ื ืืงืืื ืฉืจืื ืืืจืื ืืฉืจืื ืงืจืืืื ืืขืจืืื, ืืื ืืขืจืื ืืืฉืืืืื, ืืืืขื ืื ืืืคืืฉืืกืืื.
ืจืืฉื ืืืฉ ืืฉืื ืื, ืื ืกืืจ ืขืืืก ืฆืจืื ืืืื ืืฉืจ ืืืืฉืืช ืืฉืจืื. ืืคืกืืงื ืืงืื ืืช ืืืื ืืืฉืืง ืฉื ืืืจืื ืืขื ืืืื ื-7.10, ืืืืื ืืฉืจืืฃ ืืช ืืืืื ื, ืืื ืืืฉืืจ ืืฉืืืื.
ืืฉืชืืื ืขื ืืื ืกืช ืืืงืืื ืืืฉืื ืืืืืคืืช, ืืืคื ืฉืขื ืืืช ืงืืื!
Sahar Massachi ๐ฃ retweeted
This talk is a beautiful articulation of one of the most important design problems of our time. A few thoughts on planning with agents:
- I'm finding one of the best inputs to plans are recordings of live conversations among a team. this tends to capture a *ton* of peripheral useful context, that's hard to capture even in a solo voice recording -- because people elicit intent from each other. A generalization of this principle is that agents thrive on broad, high level, "messy" context - just like senior employees.
- I think visual outputs from agents are critical to complement text, but I'm actually of mixed minds on interaction. Yes, it's clearly nice for certain kinds of things like twiddling visual parameters. But most of the interactivity my agents produce feels mostly designed for dopamine reasons and not providing much utility. I also think interaction can inhibit deep thought compared to linear reading. I actually think for many deeper problems I would prefer top notch prose -- Richard Dawkins / Steven Pinker / Annie Dillard level -- from my agents over any kind of flashier output.
- One mental model I find consistently helpful for planning: the job is to establish a *design space* together between human and agent -- that is, a set of relevant dimensions of variation -- and then to efficiently explore the space. Getting the dimensions of the space right is actually one of the hardest parts. Once you have that, you can ask an agent to plant a bunch of starting points spread throughout the space, and then locally hillclimb from the best one.
New talk! About the divided human & agent worlds, boundary objects, and building thicker interfaces
I've been exploring how to improve "planning", aka. deciding what agents should build, how you want it built, & making the whole team part of that process
maggieappleton.com/planning-โฆ
Sahar Massachi ๐ฃ retweeted
we need to bring back company toilet engineering newsletters
facebook had one called โthe weekly pushโ
No time wasted at the @modal bathroom, where I left the urinal knowing more about DST than when I arrived
๐๐๐๐๐๐๐
The GRU hacked into Jared Kushner's iPhone. Dutch military intelligence (MIVD) then recruited a GRU spy who gave them access to the hacked phone. What MIVD discovered was that Kushner had access to two bank accounts --one Azerbaijani, the other Israeli --where Saudi Crown Prince Mohammed bin Salman and Emirati ruler Mohammed bin Zayed deposited money. It sounds like a slightly far-fetched episode of "Slow Horses," but it's real. And it proves just how eye-poppingly corrupt this administration was and continues to be as its "special envoy" decides the fate of tens of millions in Europe in the Middle East. Simply incredible reporting by Matthew Cole, drawn from his new book on Kushner. vanityfair.com/story/jared-kโฆ
Sahar Massachi ๐ฃ retweeted
CLAUDE FOUND THE SEEDS
anthropic.com/research/ecdsaโฆ
I find this half fascinating and half baffling
Gonna marinate in this one
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Everybody:
Me: SMASHing that preorder button like a looney tunes on rocket boots dunking a basketball
Oh no was I wrong?
xitter.cf/ichbinilya/status/2109โฆ
I guess we will have to let the dust settle and observe dispassionately
Google: Weโre rolling out sophisticated AI, itโs amazing, youโre welcome.
Also Google: Gmail still canโt figure out that when people email me en francais asking for โHortenseโ theyโve got the wrong guy.
SMASHING that retweet button.
The most important thing is to actually seek and speak truth.
Finding truth is really hard, actually.
And itโs easy to get wrong.
I think all journalists and academics could stand to read this essay on the limitations of "speaking truth to power."
"Intellectual courage is extremely important. But without first carefully establishing whatโs true, itโs either pointless or harmful. So intellectualsโ primary responsibility is simply to seek and speak the truth, full stop. This job description is less heroic, but itโs more intellectually and socially useful, and itโs less vulnerable to self-deception precisely because itโs less heroic."
conspicuouscognition.com/p/sโฆ
I joined Notion recently! Thereโs no place Iโd rather be. And Iโd love to tell you why. (On 7:30pm on a Friday night? Ooops)
Eighty one years ago, as the war wound down, America's head of military research, Vannevar Bush, laid out a particularly un-warlike dream in the pages of The Atlantic Monthly. The quest to realize that dream has driven computer people of all kinds ever since.
What was it? He called it the memex: a personal library of not just all the world's knowledge, but linked and sorted and indexed so that trails and hops of thought could be stored and recreated. He thought it could be created on microfilm -- microprocessors hadn't even been invented yet. But he had a vision.
Chasing this dream occupied Doug Engelbart, Alan Kay, and other computing pioneers of the 60s and 70s in Palo Alto. It stalked the halls of Stanford and Xerox, Bell Labs and Apple. Tim Berners-Lee's hyperlink was both a shadow of its grand design and a huge leap forward.
The dream is a bicycle for the mind. The spreadsheet helps you do math, but won't run your accounting for you. The google search gets you information, but wonโt write your memo. The phone tracks your friendโs birthday so you remember to call them. Computers can help people be better at being people.
But the dream has a fatal weakness. Do we want a beautiful store of information, neatly organized, ready to make all questions trivial, all code easy to write, all work a snap? Sure. But putting in the annoying work of keeping it all organized and up to date โ no thank you! Weโve made progress, but the dream of a real memex, a place where information is organized resiliently and flexibly โ weโve never done it. Till now?
Now, with LLMs, we can finally make the dream happen. You know what things are good at the repetitive knowledge work of moving information from page A to page B? Robots! But only if you use them right.
Thing is, since the 70s, we tech people have been fighting each other about the role of AI. Are computers here to replace human thinking, or to augment it? Are we building robots and factories, or are we building wikis and libraries?
Today, this battle has never seemed more relevant. Team "robots will do thinking for you" is rolling out wonders. Team memex, team "robots will help you be smarter" needs to put some more wins on the board.
And we're doing it. We're doing it at Notion.
I've spent the last year or so getting my arms around these sorts of problems. And, when you get less grand and more prosaic, much of it boils down to this:
What info (what context) do agents get when? How do you push it, and how do you help them pull it (search it)?
So that's why I joined the search and context team at Notion specifically. I like it here. Notion places a premium on "direct and kind" in communication. For an unshakeably earnest guy like me, that's perfect.
We've got some great stuff in the hopper and I look forward to telling you more soon.
Thanks to @geoffreylitt and @stevekrouse for helping me get here, and shoutout to @nbaschez @varadh @ivanhzhao @ishverduzco @iscoe @sarahmsachs and all the others who welcomed me in when I got in.
Followup: xitter.cf/sayhar/status/21087105โฆ
Why did I finally go public on Friday night? Because my team built this!
Kinda. @baseten used off-the-shelf Notion components to build this. But it took effort.
What if it took ... like, no effort? That's where we come in.
On @NotionHQ, right now, in 30 seconds, you can:
- set up a self-improving robot
- mediated via slack
- with access to all your Notion and slack data
- keeping a database of knowledge up to date
(Click on the "agents" button)
And, tbh. It's just the beginning.
For decades we've dreamed of the perfect library. We've taken so many shots at it -- the memex, the zettlecasten, wiki, the system of record, whatever. And it always falls apart because keeping it updated is annoying.
Robots are good at annoying librarian tasks. The dream of the self-healing wiki is in our sights. We can build it. Today. On Notion. But what if we made it easier and easier....
Why did I finally go public on Friday night? Because my team built this!
Kinda. @baseten used off-the-shelf Notion components to build this. But it took effort.
What if it took ... like, no effort? That's where we come in.
On @NotionHQ, right now, in 30 seconds, you can:
- set up a self-improving robot
- mediated via slack
- with access to all your Notion and slack data
- keeping a database of knowledge up to date
(Click on the "agents" button)
And, tbh. It's just the beginning.
For decades we've dreamed of the perfect library. We've taken so many shots at it -- the memex, the zettlecasten, wiki, the system of record, whatever. And it always falls apart because keeping it updated is annoying.
Robots are good at annoying librarian tasks. The dream of the self-healing wiki is in our sights. We can build it. Today. On Notion. But what if we made it easier and easier....
So, @baseten built themselves an answer factory that gets smarter over time.
Questions go into Slack. Answers post automatically. And when it doesnโt know? It learns and updates all its sources for the next time.
Try it yourself with this โจmagic link โจ app.notion.com/?target=answeโฆ
Here's the essay, btw. Courtesy of @worrydream -- a modern day keeper of the flame: worrydream.com/refs/Bush%20-โฆ
























