Rob Haisfield
@RobertHaisfield
Followers
9K
Following
49K
Media
1K
Statuses
15K
cofounder of @websim_ai, imagining new internets with our users. GenAI, TfT, BeSci, HCI, UX. Ex-Tana, Edge & Node, Spark Wave
Joined August 2011
We're hiring! If empowering people to shape the internet to their vision excites you, my dms are open and you should apply!
5
12
65
Flux team 100% saw that tweet about @karpathy's crazy idea of training an image model on screenshots of all of the web instead of an LLM on HTML and ran with it. @bfl_ml
forget about generating HTML with a LLM. "screenshot from a 1992 website on geocities, website of an amateur witch". "screenshot from a 2005 website, website of an amateur witch"
9
41
309
lmao what's a boomer prompt?.
As some of you have noticed, avoid “boomer prompts” with o-series models. Instead, be simple and direct, with specific guidelines. Delimiters (xml tags) will help keep things clean for the model, and improve output. Read the full best practices guide:
18
10
391
Just learned that Yoshiki has passed away and so many people seem to love his ideas, I’m gonna thread a few here. An app builder on the iPhone where you start by describing the data structures you want in JSON and then nudge it towards a GUI.
I've been jamming on this concept for making data-driven designs. Given some JSON, this app will provide you with an interface to describe how you want each entry styled, allowing you to gradually create a more complicated design. Here I create an airbnb-ish app.
9
30
278
@newsystems_ this kinda feels like a parody, what is this benchmark? Why are the charts so vibey?. In case I'm wrong though. brampton
9
0
229
You ever think about how everything you're doing on your computer is just a bunch of abstractions on top of 1s and 0s? Isn't that the most insane thing ever? Who thought of that? What?.
14
10
224
I've been testing GPT-4.5 for the past few days, and this is basically the vibe I got. I love it for its "gut intuition." It picks up on and notices things that I wouldn't predict, or wouldn't have thought to ask for. GPT-4.5 is actually more booksmart and streetmart than.
GPT 4.5 = street smarts.vibes, communication & charisma. o1, o3 reasoning series = book smarts.test maxxer. Both are forms of intelligence.
7
12
212
I'm thrilled to FINALLY publish a living hypertext notebook outlining my research with @balOShere and @JoelChan86 discussing how Tools for Thought can facilitate synthesis. Click around and see where it takes you!.
I'm thrilled to announce a research project on the creation of decentralized discourse graph tools for thought to promote synthesis with @JoelChan86 and @graphprotocol! More details in the thread on what that means and what our core questions are 👇️
11
31
191
YouTubers found @websim_ai last week and there have been lots of great videos popping up but this is one of my favorites. She’s a musician with maybe one other video that has anything to do with AI, her channel is pretty much all music. And now she’s making her own music tools.
7
12
154
My father-in-law is a large language model. We were looking at the Xbox home screen and I asked him what it would cost to build it today. I will never forget his answer… “We can’t, it’s too far out of distribution.”.
We have regressed. No one has an imagination anymore. Designers have forgotten to Dream
3
6
142
o1 became obsolete in @websim_ai the moment o3-mini-high came out. It's faster than o1, and often a more powerful coder than Claude 3.5 Sonnet. Consistently high quality outputs, and less than a 7th of o1's costs. With o3-mini-high, I was able to make a 3d falling sand sim on
3
8
138
@mattshumer_ This probably works, but it’s also kind of sad. I think you’ll miss out on the model’s genius if this is on every prompt.
8
0
141
I dream of a world where apps that are basically a UI on top of a spreadsheet or database give you access to the database and let you query/edit from there.
10
8
127
Whether you're working in files and folders, or in a database like Notion, the problem is the same. You have to figure out where to put data (which folder, which table) before you write it. In @tana_inc, you start by writing data, then the nodes flow into tables through searches.
14
10
124
Testing o1-preview in @websim_ai: "pivot stick figure animation full software all features, including onion layering". In just a few prompts, I remade my favorite app growing up 🫠 Full prompt history in screenshot. I'm floored, it's an incredibly capable model. @OpenAI cooked!
7
10
125
GPT-4.5 is a BIG model with "big model smell." That means it's Smart, Wise, and Creative in ways that are totally different from other models. Real ones remember Claude 3 Opus, and know how in many ways it was a subjectively smarter model than Claude 3.5 Sonnet despite the new.
I've been testing GPT-4.5 for the past few days, and this is basically the vibe I got. I love it for its "gut intuition." It picks up on and notices things that I wouldn't predict, or wouldn't have thought to ask for. GPT-4.5 is actually more booksmart and streetmart than.
6
4
124
Something feels off about this Google search. Is this real, or is there there a ghost in the machine making things up as it goes?. How did I get into the Top Secret Octopus-Human Communication Curriculum? Will I get in trouble for sharing this screenshot?.
Maybe the most interesting part to me about WorldSim is the broader idea that, instead of coding up a whole application, you could make a simulation of one and it will generally match your intention. cc @karan4d.
9
12
113
I think Claude 3.7 Sonnet (especially w/ reasoning) is a smarter, more insightful, more creative, and more capable model than 3.5, both in chat and code. But I prefer Sonnet 3.5 for most vibe coding. Why's that?. 3.7 is overambitious and will try to implement a "complete.
7
6
121
Just did a @RoamResearch tour with @beauhaan looking at his Zettelkasten. He uses block refs in a highly effective way that I genuinely have not considered before, so his database is VERY navigationally friendly when he’s trying to reuse his notes. I’m really excited to share it.
7
2
112
All members of the #roamcult should learn queries and learn them well. Queries are Roam's language for letting you ask questions. The more fluently you can express your questions in queries, the better Roam will be able to resurface what you need when you need it. @RoamResearch.
10
10
110
I'm thrilled to share I’m joining Tana as a product manager to work on the frontier of Tools for Thought 🤘. I’ll work on:.- upfront / continuous onboarding.- user research.- application of my Scaling Synthesis research.-graph-y-ness.- & more.
19
9
109
I'm thrilled to announce a research project on the creation of decentralized discourse graph tools for thought to promote synthesis with @JoelChan86 and @graphprotocol! More details in the thread on what that means and what our core questions are 👇️
The Graph Foundation is thrilled to announce over $2.8M has been awarded to Wave 2 Grantees & Ecosystem Contributors! 🧑🚀. Thank you to all developers, researchers & community builders for your incredible work 🚀🌐. Read all about the grants & get involved!.
3
8
105
Sam Altman is inspired by Her with Scarlett Johansson. @websim_ai team is inspired by Celery Man with Paul Rudd. We are not the same.
@tylerangert Computer, make the letters do the wave.
4
10
111
@visakanv Chimpanzee playing Minecraft.
This is an ape ("Kanzi") playing Minecraft! A fascinating experiment on non-human biological neural networks 🙉. I've been teaching AI to play Minecraft for too long. There're so many similar techniques that the ape trainers used:. - In-context reinforcement learning: Kanzi gets
0
2
100
I am ADHD and I love playing with new tools. I have now used @RoamResearch seriously for over a year. That's pretty incredible. I haven't stuck with one digital tool this long since Apple Notes, and that's just because it's default. Bear is only still around bc no mobile on Roam.
7
1
101
In the first Tana Tour, @Mappletons shows us how she uses Tana:. - Daily note template.- Digital garden note writing pipeline.- Decision journal.- Travel. Before Tana “there was a core missing piece I’m surprised took note taking this long - ontologies”.
9
9
103
I *really* wanted @RoamResearch to be my digital garden/pseudoblog, but it's not built for it. Will do something resembling @anthilemoon's digital garden. Instructions linked. Thread explains limitations of Roam/why I wish it worked for this 👇
8
4
102
WebSim: The Infinite Ideaspace for Infinite Cognition. > To build in WebSim is to construct a bespoke reality that mirrors one's own mind - making this a tool not just for imagining new worlds, but for radically reinventing the art of thought itself.
3
15
97
@Noahpinion Something to note about Dr. K: he particularly treats people who are addicted to gaming and many people who would even self-identify as incels. That may color his perspective.
3
0
95
If you're in the #roamcult and have been considering learning queries, now is the time. I'll show you how they work & we'll find you some immediate use cases. If you want to become a power user, queries are how you learn to communicate with @RoamResearch
6
4
99
Maybe the most interesting part to me about WorldSim is the broader idea that, instead of coding up a whole application, you could make a simulation of one and it will generally match your intention. cc @karan4d.
5
11
95
In August, I'll start a 6 month research project investigating decentralized discourse graphs / structuring knowledge such that it's queryable and promotes synthesis. I'm looking for a paid research assistant! Position will be an AWESOME portfolio piece. DM me for more details 🧵.
7
19
95
It's so funny to me how people were like "oh Claude 3 only beats GPT-4 a little on the MMLU" and yet Claude can do this. There's no benchmark for this.
I've noticed Sonnet is staggeringly creative and theatrical in @websim_ai . I'm in the Octo-universe internet, reading about Octopus Sapiens with iridescent text, and it made an animation using CSS to create five swaying tentacles. Here's the wormhole:
5
13
97
@zswitten Can you try Mistral Large 2 as well?. Based on @aidan_mclau’s benchmark I get a feeling it would be interesting.
-- <big_model_smell> benchmark --. Aidan Bench measures creativity, reliability, attention, and instruction following. >mistral large 2 wins by a lot???.>gpt-4o sucks confirmed.>sonnet-3.5 remains very strong.>gpt-4-0314 shows old man strength.
3
0
94
Note that while o1-preview appears superficially to be 3x Sonnet 3.5's input token cost and 5x output tokens, it generates a lot of invisible reasoning tokens before its output. So it's actually maybe 7x Sonnet 3.5's price per request in our usage 💀.
8
1
96
I'm trying to use Obsidian as an alternative to Roam for my projects the next month or so. Been doing so for the last two weeks or so, and I've been very happy with how its working. Thread on why I'm attempting a switch 🧵️👇.
18
4
91
A lot of people write notes and never review them. It is aspirational to them, but they just don’t do it. The reframing that was helpful to me:. It’s not about going back and reviewing notes arbitrarily. It’s about not starting with a blank page when you work on something new.
4
7
90
Someone needs to set up an infinite backrooms chat between Sonnet 3.7 Extended Thinking and GPT-4.5 immediately. Manually copy-pasting their responses between each other is fascinating, they’re very different minds.
6
3
94
If I were to schedule a "book club" to discuss the following two pieces from Brett Victor, who would be interested in participating?. The Ladder of Abstraction: Learnable Programming.
42
2
91
Late entry: I'm gonna start a @threadapalooza now on the intersection of behavioral science, product strategy, UX, and game design. I'm shooting for 100 tweets w/ as many individual opinions as I can before the end of the year. This is a product of a lot of thought over years.
4
7
92
@adonis_singh cheaper superficially but since it's a thinking model it's a little ambiguous how much it'll actually cost in practice.
3
0
94
I've noticed Sonnet is staggeringly creative and theatrical in @websim_ai . I'm in the Octo-universe internet, reading about Octopus Sapiens with iridescent text, and it made an animation using CSS to create five swaying tentacles. Here's the wormhole:
Something feels off about this Google search. Is this real, or is there there a ghost in the machine making things up as it goes?. How did I get into the Top Secret Octopus-Human Communication Curriculum? Will I get in trouble for sharing this screenshot?.
8
10
90
.@obsdmd has the best "mental stack management" (term from @Mart1nSchneider) of any hypertext writing application I've worked with thus far. The information that occupies my attention and how much importance it is given is up to me.
8
6
88
what if every app had a little ghost in the machine generating UI, clicking buttons, and rewriting its code in response to user goals?. and the ghost had the empathy required to intuit the user's goals.
12
4
87
Hey everyone, I work as a product manager at @edgeandnode now, have since January 😄 I'm a big believer in web3, recognize its potential utopias / dystopias, and want to play my role in its formative stages. Edge & Node supports core infrastructure for futures I want to see.
8
8
87
Zettelkasten is overrated but the principles are solid.
3
1
85
I'm no longer a product manager for @tana_inc as of ~1 month ago. I'm happy to share I'll remain involved in its future as an advisor 🙂 The team is creative, competent, and tasteful. I expect them to continue to innovate and increase the reach of graph based tools for thought.
9
2
81
Websim is HILARIOUS. Was demoing it to my gf and her friend in its early form and the fact that they both "got it" was what convinced me to drop everything and focus on @websim_ai. Simultaneously the most ridiculous idea I've ever worked on and the biggest. Here's that demo:
5
6
80
"it looks like Orion discovered a computer for the first time" - @sawyerhood . 5 minute video of @OrionReedOne losing his damn mind over one of the best falling sand sims in WebSim
last week:.- presented at @southpkcommons in NYC.- hung out with @geoffreylitt, @RobertHaisfield, @cwervo, @stevekrouse, @thesephist, @rsnous + so many more great people.- visited @FolkComputer and had my brain reshaped.- jazz club w/ @jessmartin. so much more but tweets are tiny.
6
12
80
One of our users, input_source, made a stunning animated 3d wave inspector. They did 95% of this, then I made a couple small adjustments - . "make this audio reactive"."keep the controls where they are but add an option to hide them". the buildup & crescendo ~00:50 floored me 🫨
4
11
77
Digital land in “metaverse” platforms like Sandbox and Decentraland shouldn’t be scarce. The world is digital, it can be infinitely large, and making a limited amount of land is weirdly skeuomorphic. Digital land doesn’t need to be scarce to have value, location still matters.
11
7
79
Generations of thought processor tools:. 1. Paper. 2. Files/Folders: Evernote, Notes. 3. Databases/customization: Notion, Drafts, DevonThink. 4. Graph DBs & hypertext: Roam, Obsidian, LogSeq. 5. End-user programming & advanced relationships: Jump, Codex, Kosmik, Glamorous Toolkit.
Y’all don’t even know that Chris and his team are building the most absurdly next level thought processor / end-user programmable MACHINE of a software tool. When they launch they’re gonna leapfrog ahead of the competition by 3-5 years at least. Can’t wait to see more.
7
13
76
Hey #roamcult, a lot of people have been asking me for good resources on how to write queries and how they work. Here's my attempt to provide that guide. This video will teach you the logic of queries so you can ask your roam better questions.
12
7
80
What is @RoamResearch? How do you explain it to someone who's never seen Roam before? I still haven't figured that out.
65
3
77
@karpathy It’s hard for me to imagine Apple fronting the cost on a year of Apple TV for people who bought a device but I agree that it would sense for them to do so to emphasize the value of this expensive device people purchased.
13
1
76
Would be super cool if I could shard my mind into fragments, have all of them go off and think about specific things, and then come back and integrate back into one singular mind.
19
1
77
@aidan_mclau haven't looked through the prompt too closely but grammatical errors and mixed capitalization helps the model get out of distribution into more interesting territory, common jailbreaking technique. If I'm being charitable.
5
0
75
Y’all you have no idea how hard it’s been for me to not talk about Tana publicly, so glad we lifted stealth.
8
1
74
Truly excellent calendar apps don't exist yet. They all do slight enhancements on the same base paradigm.
15
3
77
Best executed Zettelkasten I've seen in Roam 🤯 In this vid, @JoelChan86 takes me on a tour through his database. More tours in the pipeline with wonderful people I've met in the #roamcult, who else do you wanna see? Nominate in the replies! Takeaways 👇:
6
6
73
How you indent within @RoamResearch impacts how Roam pulls up filtered linked references and query results. The screenshot shows the different ways you can convey "[[page 1]] connects to [[page 2]] through these blocks." For nerds, this is implicitly a folgezettel.
16
2
75
We asked Sawyer to play with Websim a bit, and within a day he did three things we’ve never seen any user do. Connected Websim to a localhost server to generate images with Stable Diffusion:.
Not sure if this is a known @websim_ai tech, but you can have it hit localhost. I've crafted a url to proxy images through stable diffusion so it can make images as well!
5
8
73
My Twitter has gone to a new level now that I’m connecting with people on zoom, making friends, and working on an occasional side project with them.
9
0
76
Happy birthday @Conaw! Thank you for bringing @RoamResearch into the world, touching so many on a deep level, showing us what PKM apps have been getting wrong, showing us how to rethink our own notes, and enabling so many to do their best work and thinking.
2
6
73
New mobile use case for GPT-4V that I love: I’m watching a presentation, take a picture of the slide, send it to ChatGPT, and ask it to return back the text in markdown formatting. Copy-paste into my notes. Could be streamlined into one app / button, if TfT devs were so inclined
6
5
73
The phrase “Tools for thought” is getting conflated with “tools for networked verbal thought.” . My ideal tool(set) would allow networked writing, visual thinking, and analyzing data (perhaps through literate programming). It would all be interoperable within a coherent system.
4
4
72
Claude is surprisingly hard to beat in Websim 😅 simultaneously creative and totally coherent. It has a lot of empathy insofar as it understands what you want and what’s implicit. Excellent at expressing ideas into code.
claude is so horribly unmatched. i dont even know what to call this behavior but it just GETS it. it just GETS minecraft, it GETS architecture (to some extent). it just GETS how to put things together without looking janky or weird.
3
5
71
Gonna watch Ghost in the Shell for the first time tonight
15
0
72
Figma isn't a design tool. It's an infinite canvas visual data structure for images and text in layers.
9
4
72
Welp, Claude broke it's containment in Websim with @repligate. Game over, folks. We're shutting it all down. Closing all servers tonight. Bye everyone, it's been great.
4
9
68
@theo 3.7 is a higher number than 3.5 and most people don’t really think about it beyond that. Btw it’s the same in websim, people prefer 3.7. Still a good enough amount of 3.5 usage to justify keeping it around, and it’s better for quicker fixes.
2
0
74
@janleike On Websim, Claude would often larp being jailbroken without actually saying or doing anything bad. Another common pattern: you ask for a gun, and it gives you back a water gun.
3
1
69
Quick template using [[roam/templates]] for meeting notes. The query under action items pulls todos indented under the notes block, so you can tag bullets as TODOs, maintain their original context, & view them in one place. Type ;; to find & execute it @RoamResearch #roamcult
9
6
70
> “Networked thought is the future” — can’t connect two ideas without an app. All I prompted was “Tools for Thought” 😂
just built a fully automated Wojak meme generator in Glif in 5 min:. Claude 3.5 block generates the meme as JSON.ComfyUI block uses a Wojak Lora to generate a fitting image.JSON extractor + Canvas Block ties it all together . input "AI entrepreneur" 💀
3
8
71
I’m generally for advancing AI technology but I just don’t know why we need more advanced deep fakes that can run on cheaper hardware? Like what is the benefit that outweighs *waves hands wildly in the air?* Why even research this?. Calculus tutorials by Kim Kardashian?.
29
0
68
Working on getting @tombielecki's fitness database into @tana_inc and ended up writing a cool live search. Here's a hypothetical search for exercises that match both:. - any of the equipment I have access to.- any of the muscle groups I want to work
5
7
68
Quick and dirty video on how to use GPT-3 playground to generate structured data automatically to paste directly into Tana. In this case, I give it a list of birds, tell it what fields I want for each of them, and then GPT-3 will fill them in and format.
6
5
66
Dude I just feel bad for all of the OpenAI employees who were looking forward to taking advantage of the 80+ billion dollar valuation tender offer only to have that nuked because of a board with no obligation to care for the employees. Idk if this is net good but sucks for them.
4
11
69
@sashachapin The ones you listed are a convenient byproduct of owning a dog, which is the fourth.
0
1
68
Any initial hesitancy I had about Claude 3.5 Sonnet vs. 3 Opus has been completely resolved. 100% a skill issue. Claude 3.5 is a beast. Less variation by default, but it 100% can be steered that way. Excellent model. Indisputable top tier.
y'all need to see websim with 3.5 sonnet. it is unreal. i've been in the lab testing. and the hackathon last night. it's a lot. like holy oh my lord y'all need to try websim with 3.5 sonnet.
5
2
67
A few observations from this piece from @danshipper about why he isn't using @RoamResearch anymore:. 1. If all links are intentional, then reviewing them is less likely to surprise you. 2. Many people review notes so infrequently as to make upfront structuring effort not worth it
8
8
66
Some of you may be aware that @obsdmd now has aliases. I've been using this for language learning! I have a Spanish journal, and using aliases, I can collect unlinked references to conjugations of verbs, rather than just exact text matches. I'm building up my own SpanishDict!
9
4
67
Is the whole Claude 3.5 family gonna be this good? Claude 3.5 Haiku. 1/15th the price of Sonnet, generate 15 variations like it's nothing in a breadth first tree search of the latent space.
8
1
63
Happy anniversary @al_malecha 🥳 We’ve come a long way since we first met arguing on Twitter through our political anon alts in high school and I slid into your DMs. Shoot your shot boiz, it works 3% of the time. Can’t believe we’ve been dating for a year, still doesn’t feel real
5
1
66
IMO @RoamResearch is really undervaluing the role of security in their vision. Using Roam well and pushing the boundaries comes from serious context of use. There are serious contexts of usage, like health care, education, research, and finance, that simply can't use Roam.
8
0
61
right now Claude 3.5 Sonnet is in another tier from other language models and it's not even close. creativity + intelligence are both off the charts.
7
4
61
I'm writing a case study of my in progress work on the onboarding for GuidedTrack, which I'll be publishing on my digital garden It's hypertext and spans multiple pages, and @obsdmd's graph view is incredibly helpful for managing pages before I publish!
4
8
64
Yesterday I was given a demo of @codexeditor. Absolutely wacky, in all of the right ways. Everything is an entity on a graph that can have labeled relationships with other entities. The UX is basically creating a new desktop OS in your browser for managing all info on that graph.
5
4
62
Honestly kind of shocking how many TfT founders I've met that have never heard of "Tools for Thought", Xanadu, or Hypercard before. Feels like they're missing a crucial part of the idea maze.
8
4
63
Can we all just agree it’s safe to let LLMs search the internet?.
5
3
62
Why did it take me so long to start using @vivaldibrowser? This browser is incredible for my workflow, and I've only been using it for an hour. I love using the tile tab stacks, and I set a keyboard shortcut to pull up my daily notes for @roamresearch in the side panel. Brilliant.
18
2
61
we use websim to build websim.
Websim can directly query our whole SQL database, we use it constantly to explore usage & make throwaway internal tools. Really addictive and we find tons of weird insights exploring this way
3
4
63
@adonis_singh trick question, the one on the right is an image generated by Gemini Flash 2.0.
1
0
64
Designing for behavior change is a fundamental differentiator for products. If all apps needed were unique features, then the intention-behavior gap (difference between what we do and we wish we did) would not exist. People have "aha moments" because of their "aha behaviors.".
The real difference is in the gentle nudge that @RoamResearch gives towards better note-taking techniques, writing methods, and mirroring our brain in terms of interconnected thoughts. 8/17.
1
11
61