| Apr 26 |
This is next level trust building and more people should do it. The lack of humility and honesty about shortcomings with all the new AI products is exhausting and leads to constant disappointment. The best AI powered products aren’t the ones that everyone’s talking about
|
1 |
0 |
0 |
217 |
271 |
. |
| Apr 24 |
Had my first scam experience on @Airbnb today. Arrived to hotel and they had no record of my booking. Some dude just listed the place and took my cash it seems. Didn’t know that could happen on Airbnb but today I learned
|
1 |
0 |
5 |
382 |
220 |
. |
| Apr 24 |
@bentossell That tweet’s targeted at people who write code but end up creating long term unmaintainable codebases because they add libraries and complexity that compounds, which they wouldn’t if they understood what they were doing
|
0 |
0 |
0 |
0 |
231 |
. |
| Apr 23 |
@adamwathan Not sure I follow
- Button in form - visually obvious I'm in a form because I've already input form data
- <a>s that looks like a button - I can cmd+click
What are other scenarios where a form button would appear that could be mistaken for a link button?
|
2 |
0 |
0 |
0 |
275 |
. |
| Apr 23 |
YES. A lot of the newer libraries and the hip products (cough @linear) use cursor: default and it feels off.
|
0 |
0 |
0 |
224 |
108 |
. |
| Apr 23 |
Thing I've used @usescoutai to do in the last 3 days
- Send multiple emails ✅
- Send twitter DMs ✅
- Send whatsapps ✅
- Check my Linkedin DMs and send a calendar invite ✅
- Attempt to book Uber eats (got stuck)
Slowly iterating through the brittle parts but very excited
|
3 |
0 |
0 |
153 |
271 |
. |
| Apr 22 |
Made a thread 4 yrs ago w/ products I liked which had stackblitz/@boltdotnew (rocketship), @replit ($3b company), & @supabase ($2b company). All small at the time.
Good investmt thesis in hindsight: Technical founders with taste using hard tech to make complex things simple.
|
2 |
0 |
0 |
284 |
280 |
. |
| Apr 20 |
"How to build an agent" (https://t.co/H98pbsdN7S) + iterating to get this prompt + high velocity eng team + really good UX owner/s who are also technical + good GTM with a few strokes of luck = potential $3,000,000,000 acquisition. Easy to think "there's no magic tech" but actually the chances of getting all these things right is extremely rare
|
1 |
0 |
0 |
368 |
346 |
. |
| Apr 20 |
Tried O3 ChatGPT - it's really good
What it shows is that we don't need super intelligence to have super utility - it's genius is in how well it can gather, summarise, and present data, a.k.a how seamlessly the model, the UI, and tool use are integrated.
|
0 |
0 |
0 |
226 |
255 |
. |
| Apr 19 |
Every now and then I try to find a password manager I don't hate, and every time I'm disappointed. Tempted to build one - requirements:
- Don't be slow
- Don't ask me to keep logging in with no way to dial down
- Use the chrome sidebar API so I don't lose what I was doing.
|
0 |
0 |
0 |
204 |
273 |
. |
| Apr 17 |
By far the most promising tool so far has been @codegen - integrates with Slack, lets you tag it, supports threading and follow ups. The out-of-the-box code reviews were much higher quality than @coderabbitai
Major gap is it doesn't support multiple github orgs through Slack
|
1 |
0 |
1 |
242 |
277 |
. |
| Apr 17 |
Have been experimenting with more AI dev tools recently
@DevinAI: Setup was extremely promising - they had me provision a cloud VM & install the packages it needs, but the first task I gave it - it didn't actually test the code changes against the live app & the PR was bad 🤷♂️
|
0 |
0 |
0 |
399 |
288 |
. |
| Apr 17 |
@AravSrinivas Question: Why build this as a new browser vs a cloud browser that someone opens once & logs into their services, then interacts with via tasks/chat to ask it to do stuff?
|
0 |
0 |
0 |
0 |
188 |
. |
| Apr 17 |
A big part of what's allowed us to scale @rlygoodsoftware is settling on tools and sticking to them - @heroku is still our go-to, we're amassing quite a collection. https://t.co/VSFFgwMDvA
|
2 |
1 |
1 |
266 |
188 |
. |
| Apr 16 |
I'm looking to purchase a last minute @ReactMiamiConf ticket, potentially from someone who has one but can't make it. DMs open 😁
|
1 |
0 |
0 |
184 |
128 |
. |
| Apr 16 |
@kraustifer @ReactMiamiConf Hey! Are you selling?
|
0 |
0 |
0 |
0 |
49 |
. |
| Apr 14 |
@scrimba Any plans for ruby?
|
1 |
0 |
0 |
0 |
28 |
. |
| Apr 13 |
True. Now apply this to software engineering
|
0 |
0 |
0 |
264 |
44 |
. |
| Apr 11 |
@marcoroth_ @rubykaigi @railsconf Oooh please share some sneak peaks this sounds suuuper interesting
|
2 |
0 |
0 |
0 |
100 |
. |
| Apr 10 |
Every time I think about this tweet I cringe. I was so wrong about this. Azure has a slightly nicer UI than AWS but the cloud platforms are so user-hostile it's crazy https://t.co/fBshRhWJoQ
|
0 |
0 |
1 |
271 |
190 |
. |
| Apr 09 |
@aviflombaum @michelson Avi have you seen this? I'd guess we're 99.9% philosophically aligned on these matters generally. My opinion was changed recently only when I realized 1. The web component searchable-select I built is reeally hard to make work performantly, and 2. That I can drop in preact with…
|
0 |
0 |
0 |
0 |
303 |
. |
| Apr 09 |
@aviflombaum @michelson Sure! Here was one we had recently:
- 5000 contacts with first name, last name, phone searchable with plain text search
- All contacts loaded on page load - no fetching from server on keydown
- Each item also has nested state (invited=true/false can be toggled) with state… https://t.co/kOZTcDStiW
|
0 |
0 |
0 |
0 |
321 |
. |
| Apr 09 |
@aviflombaum @michelson I’ve evolved on this. I agree for a hello world div it’s dumb but Try to build a performant 5000 item contact picker in
- vanilla js
- web components
- react
That’s when you do realise that Jsx as a templating language is best. I still avoid builds and write mainly html but…
|
0 |
0 |
0 |
0 |
301 |
. |
| Apr 09 |
@staysaasy Sounds like ADHD potentially being the common thread there?
|
1 |
0 |
0 |
0 |
70 |
. |
| Apr 08 |
A lot of tech people have adopted a mantra of “You can just say things” when it comes to AI. I don’t think they realise that saying lots of things and either not caveating or not correcting as they learn more, hurts their credibility among smart circles.
|
1 |
0 |
0 |
175 |
254 |
. |
| Apr 08 |
This pattern is objectively better for many things we currently use APIs for - not sure why it's not more popular.
|
1 |
0 |
0 |
184 |
114 |
. |
| Apr 08 |
Some patterns I've noticed
- Hiring developers is still incredibly slow & not a solved problem.
- CVs which just list languages & libraries the candidate has worked with were always silly, but are even dumber in the age of AI assisted coding when smart devs can learn new tools in a day/week.
- The correlation between experience and skill is becoming weaker. People with 2 years experience can have the skill & output on par with ppl with 5+ years.
- Vice versa, people with 5 to 10 years experience can have less output than many juniors because they've been at companies with norms of extremely low velocity.
- Communication (written & verbal) & time management are as-if-not-more important than technical skills but are difficult to screen for.
- It's basically-impossible to predict how someone will perform before you've seen them produce something.
- As a hirer, sifting through lists of dozens of CVs is a huge pain and gives very little confidence you're actually getting the best.
One thing I would love to see experimented for dev hiring is video - some way for devs to record videos of themselves solving problems. Pair this with a multi modal AI and I think you'd be able to extract a very accurate rating very quickly, meaning the hirer wouldn't have to watch hours of video and could just skip to the good parts. Anyone know of anyone doing this?
|
2 |
0 |
1 |
230 |
1.4k |
. |
| Mar 31 |
@ctoincanada @rails Jsx just has far better default conventions for state management and templating for small components. Simple example recently was building a client side contact search over 2k records. Using the “sprinkles” approach with dom to store state gets slow fast, using…
|
0 |
0 |
0 |
0 |
282 |
. |
| Mar 31 |
@ctoincanada @rails I don’t see them as being mutually exclusive. You can use Turbo and have little islands of react there too - it’s stimulus that this would replace
|
0 |
0 |
0 |
0 |
166 |
. |
| Mar 31 |
My new favourite pattern for using islands of React in a @rails app (or any app for that matter). No build step, super lightweight, readable & maintainable. Perfect if you want to use React for little bits without having to replace the whole UI. https://t.co/v8U4BUvyCs
|
15 |
0 |
2 |
1.1k |
273 |
. |
| Mar 30 |
@Ryan___Doyle Did you try browser use cloud? What was the blocker there? (We're building a similar product but not fully production ready yet)
|
0 |
0 |
0 |
0 |
142 |
. |
| Mar 29 |
@robotiguy Custom project in Claude chat that I've given explicit instructions to only output component code - so I can always one-click copy the entire component. I'm working on a component library so the only context it needs is in the file - doesn't need to know about anything else
|
1 |
0 |
0 |
0 |
286 |
. |
| Mar 29 |
Just realised. I've been working on a few UI components for several hours - copy-pasting the entire file into Claude repeatedly because it's faster than using cursor, and there hasn't been a single instance where it's returned broken code or added unwanted code out of 50+ passes
|
4 |
0 |
2 |
905 |
279 |
. |
| Mar 28 |
Something happened Claude since yesterday. 3.7 is now one-shotting things it was previously failing completely at. It's so good
|
3 |
0 |
0 |
1.3k |
127 |
. |
| Mar 28 |
Did you know you can have the niceties of React's templating language with no build step and a mere 6.7kb import?
|
3 |
0 |
1 |
488 |
113 |
. |
| Mar 27 |
I would pay very good money for a codepen with AI in the editor. Other options just feel to heavyweight. Does this exist?
|
0 |
0 |
0 |
225 |
121 |
. |
| Mar 26 |
Online banks are wrapper businesses
|
1 |
0 |
0 |
126 |
35 |
. |
| Mar 26 |
Biggest update to my productivity recently is @foragemail - simple idea, well executed
- Sign up with gmail
- Now all transactional emails skip your inbox
- One email per day with a digest of everything
I'm actually responding to emails again. Highly recommended.
|
3 |
1 |
1 |
583 |
265 |
. |
| Mar 26 |
I keep seeing people talking about prompting, then "going away for a bit" and coming back to it. I struggle to believe the total time taken for these people to build things is lower than if they simply wrote the code themselves.
|
1 |
0 |
0 |
178 |
228 |
. |
| Mar 26 |
Great example of current AI promise vs reality
The pressure to ride the hype train is so strong (MCP most recently), but if you don't dogfood your product and take time to iron out edge case you're going in the AI slop pile in the minds of consumers
https://t.co/LfL0M9PiGq
|
1 |
1 |
0 |
225 |
277 |
. |
| Mar 21 |
The most promising thing happening in the js ecosystem is what @jarredsumner is doing with bun
|
3 |
0 |
0 |
286 |
94 |
. |
| Mar 21 |
It feels like the quality bar for new digital products is in decline. I like to try new stuff, I think at this point I'm averaging one product a week where I have to reach out to support to just get the first run to work. Mostly AI first products which are harder, but still.
|
3 |
0 |
2 |
908 |
275 |
. |
| Mar 21 |
@codegen @codegen this *looks* extremely promising, but your support for multiple orgs (adding one org later on) doesn't work. Just spent 30 minutes on it - completed all steps, approved Github org, added to Slack etc. But the slack bot is still saying it can't access the new github org…
|
1 |
0 |
0 |
0 |
288 |
. |
| Mar 21 |
@chriscoyier any plans of adding ai to the codepen editor?
|
0 |
0 |
0 |
0 |
58 |
. |
| Mar 20 |
Crazy seeing tweets from a year ago and seeing how (little) progress has been made since then (AI code review tools are still so-so nearly a year later).
|
1 |
0 |
0 |
249 |
153 |
. |
| Mar 19 |
@bradgessler Would challenge the normative statement "is insanely awkward". It's a preferences thing. You don't like it, which is fine, but that's not true for everyone. At our company we would avoid patterns like arrow 2 and 3 - they're issues with Rails best practices IMO, not erb itself.
|
2 |
1 |
0 |
0 |
291 |
. |
| Mar 18 |
@DetachlessUI hey guys! Is Detachless maintained? Trying to use it but getting a blank page
|
1 |
0 |
0 |
0 |
91 |
. |
| Mar 17 |
Anyone know why @rails uses the CLI approach for migrations by default, versus what Ridgepole enables with a Schemafile? The Ridgepole was is much nicer, but I'm wary of adopting as it's foreign to new developers.
|
0 |
0 |
0 |
210 |
213 |
. |
| Mar 14 |
@skryukov_dev Hey @skryukov_dev just found this, a little late 😅. Are you aware of any service/playground that lets you run rails (not just ruby) in the browser? Great work on this, it's awesome!
|
1 |
0 |
0 |
0 |
196 |
. |
| Mar 12 |
@kingsley_kelly How does “coding isn’t software development” address the idea that 100% of code being written in a year is ludicrous, or is that what we’re still talking about?
|
0 |
0 |
0 |
0 |
176 |
. |
| Mar 12 |
@kingsley_kelly 1. I don’t think that’s the last 1% - closer to 20%
2. Dario said “100% in one year”. How does that happen if models can’t actually reason?
|
1 |
0 |
0 |
0 |
155 |
. |
| Mar 12 |
@kingsley_kelly Literally used Claude Code the other day to do something atypical that there wasn't a lot of docs for (building a react native component that combines react native webview and react native stacker). I was so excited - first command required changing multiple files - I confirmed…
|
0 |
0 |
0 |
0 |
295 |
. |
| Mar 11 |
Announcing Scout: AI that can do more than just chat.
A few months ago a major milestone passed (somewhat) quietly: AIs can now competently control a web browser & reliably complete multi step tasks.
Now that the tech is there, there are a few products in the space, but we're still in the extremely early stages of figuring out what's possible and building the right application layer.
The best products will come from 1. Great product taste, 2. The ability to combine and deploy the existing technologies (models, libraries), and 3. The ability to ship-iterate-improve to find the right patterns & UI paradigms.
We've built a team at @rlygoodsoftware that has all of those characteristics, so rather than wait for someone else to build the product we want, we've begun working on something I'm super excited about.
Some things we're focusing on:
- Task Focussed (as opposed to chat-centric): With Scout you can quickly fire a task off or run it on a schedule - you don't need to monitor and respond to a chat (but Scout will ping you if it needs you).
- Credential Management as a first class citizen: No pasting your password in plain text into a chat. Scout has a built-in password manager that lets you save and pass your credentials safely to the Agent. We also have plans to let you deploy on your own cloud.
- Fine grained Control: You can set Scout to be Cautious, Careful, or Confident. In cautious mode, Scout won't proceed without asking for signoff on any PQAs (Potentially Consequential Actions) like making payments or sending messages. Confident mode - if you let it - can buy your groceries and have them delivered without your input.
These are just some of the things we're exploring. We've already built a V1 and are testing internally for now, but would like to get more people using the product from April. I'll also be sharing some sneak peeks on here as we build it out
If you're interested head over and sign up to the waitlist - link in next tweet.
|
14 |
0 |
1 |
692 |
2k |
. |
| Mar 11 |
There is a *lot* that is possible that hasn’t yet been done, by simply building really good UI/UX and combining existing capabilities
|
0 |
0 |
0 |
201 |
133 |
. |
| Mar 09 |
@thekitze This take is roughly as smooth brained as the people you’re mocking. The interesting questions are if and when the plateauing starts (or if it already has). A newborn baby doubles their weight in six months, to use your logic they would weigh the same as an elephant by age 6.
|
1 |
0 |
0 |
0 |
287 |
. |
| Mar 09 |
@hernansartorio @pagyco Also there's something strange happening when I try to use the "Plus Jakarta Sans" font specifically - it worked then stopped working, all other fonts are fine
|
1 |
0 |
0 |
0 |
183 |
. |
| Mar 08 |
Have tried all the landing page builders and @pagyco is still the best for quickly throwing something together. So easy to use, great work @hernansartorio 👏
|
8 |
3 |
1 |
686 |
156 |
. |
| Mar 06 |
Are there any world class design engineers out there who are html first? (E.g People like @shadcn who make incredible experiences and libraries without the react depedency)?
|
2 |
0 |
1 |
341 |
173 |
. |
| Mar 06 |
I *wish* i was recording how often these kinds of confident hallucinations happen because my gut tells me it has *increased* in the past year. That is that quality is going backwards - this kind of thing happens me almost daily
|
0 |
0 |
0 |
210 |
227 |
. |
| Mar 03 |
This might make me look very silly in the future, but I'm going to put it out there anyway: There is a reasonable chance that the quality of AI coding either will plateau, or already has.
The best coding model - Claude 3.5, was released 9 months ago. 3.7 is a step backwards. And 3.5 is still not capable of "reasoning" in any meaningful way, and in particular with patterns it has never seen before.
Yes - it's an incredibly useful accompanying tool for an engineer - I use it every day and it saves huge amounts of time. But the idea that even a mid level engineer can/will be replaced by AI - at least with the current capabilities & the trajectory of the last year - I think needs to be re-examined.
|
3 |
0 |
0 |
258 |
707 |
. |
| Feb 28 |
@thomasauros 100% - it’s just the debug-ability (or lack) that kills me. Coming from rails where if something goes wrong it’s generally trivial to figure out what it is. Not so by default in react land
|
1 |
0 |
0 |
0 |
201 |
. |
| Feb 28 |
How can anyone work like this - an hour lost to a literal react native webview wrapper 🤯 https://t.co/Q1ucEZK1uT
|
0 |
0 |
1 |
117 |
112 |
. |
| Feb 28 |
This was my experience too. If you’ve become really good at establishing and reusing patterns, it’s mostly faster to just write the code than to prompt it - makes me really wonder how inefficient most people generally are at programming and how much boilerplate is being written
|
3 |
0 |
0 |
185 |
278 |
. |
| Feb 28 |
Another week of js-ecosystem work.
Much more confident that if the industry stays building on react/npm, the need for developers won't be going away any time soon.
Just so. much. unnecessary complexity.
|
0 |
0 |
2 |
236 |
206 |
. |
| Feb 28 |
@coreyhainesco If it’s a simple app that used to be fast it’s most likely to be n+1 queries. Skip the profiling tools and Install the bullet gem, have it show you where to add .includes to your code to reduce the number of database calls
|
2 |
0 |
0 |
0 |
237 |
. |
| Feb 26 |
Yes! Now let’s do buttons, tab bars, modals, tables and a few more where applying one outer class with inheritance is better than individual SPCs 😁
|
1 |
0 |
0 |
228 |
147 |
. |
| Feb 25 |
Some days I feel like time’s nearly up for software devs. Other days i see a company that’s designed to obsolesce software developers, raising money to keep hiring them. I suppose that’s a promising signal
|
5 |
0 |
0 |
247 |
205 |
. |
| Feb 25 |
@TellaHQ hey guys, love your product. Recently every time I open it to record there's an update dialog which breaks my flow, and this just happened after a long recording, which is super frustrating. Meant as genuine feedback - I'd want to know if it was my product https://t.co/NFJfwtyFYh
|
0 |
0 |
0 |
0 |
289 |
. |
| Feb 24 |
Felt like I was quite unproductive in the middle of the day but didn't expect it to be this pronounced. @rescuetime is great https://t.co/zleSgOeWxO
|
4 |
1 |
0 |
403 |
148 |
. |
| Feb 24 |
As someone who has opened entities in Asia, Europe and the US, generally speaking the logistical price-of-entry to run a business is very high, regardless of country.
People pointing to marginal difference in ease-of-opening-a-business are missing the cost of *ongoing* requirements in accountants, bookkeepers, compliance.
Yes EU is bad for startups and US generally has fewer regulations.
But for the average business, the tax-on-entrepreneurship of not having a zero-cost way, provided by the government, to stay compliant, is what we should be complaining about, in *every* country.
|
2 |
1 |
1 |
734 |
592 |
. |
| Feb 23 |
@kingsley_kelly Yeah - just a case of unrealistic expectations - I wanted to be able to open a seldom-modified codebase and make some minor updates. I have cursor rules for my rails stuff but spending a day on that alone here makes zero sense. I spent 45 minutes prompting a very minor refactor…
|
0 |
0 |
0 |
0 |
295 |
. |
| Feb 23 |
Tried "vibe coding" today in a react (native) codebase & it thoroughly sucked the enjoyment out of the process, even using voice as input. The loop of prompting, then waiting, correcting mistakes, reloading, re-prompting, is just, not very fun.
|
2 |
0 |
1 |
255 |
249 |
. |
| Feb 23 |
Tried "vibe coding" today in a react (native) codebase & it thoroughly sucked the enjoyment out of the process, even using voice as input. The loo[ of prompting, then waiting, correcting mistakes, reloading, re-prompting, is just, not very fun.
|
0 |
0 |
0 |
0 |
249 |
. |
| Feb 23 |
@tnorthcutt @AlecStapp @Sam__Enright Basically it assumes you have 4+ person teams of highly paid specialists, which probably gets the best outcomes but also assumes no salary constraints. Someone needs to write the 37 signals version
|
0 |
0 |
0 |
0 |
234 |
. |
| Feb 22 |
I have a bad habit of making things (products, libraries, essays) and moving on to the next thing without sharing them, usually because I don't feel they're polished enough.
I want to get better at sharing this year - consider this my public accountability post. https://t.co/oFbz7q64zC
|
2 |
0 |
1 |
163 |
287 |
. |
| Feb 22 |
The products I use that are highest value per dollar spent:
@ContaboCom & @Hetzner_Online
@hatchboxio
Cloudron
Just an insane amount of value
|
2 |
0 |
2 |
212 |
148 |
. |
| Feb 21 |
I had a tweet a few years ago where I signed up for azure and was pleasantly surprised and half the Microsoft azure team liked it, then a month later I realized literally every other Microsoft software product is insanely bad and now I cringe when I see that tweet. Similar vibes here
|
0 |
0 |
0 |
317 |
284 |
. |
| Feb 21 |
@levelsio Yes but it’s react world - https://t.co/6REWvpe7q3
|
1 |
0 |
0 |
0 |
60 |
. |
| Feb 21 |
@AlecStapp @Sam__Enright This book was a zero interest rate phenomenon https://t.co/DvtTNxHH2E
|
10 |
0 |
0 |
0 |
94 |
. |
| Feb 19 |
It really bums me out that the new AI codegen tools are so far abstracted from the vanilla web that basic things like importing a css file are either not possible or have to be hacked around. Would love to see an html-first version of @lovable_dev or Bolt
|
5 |
0 |
3 |
339 |
252 |
. |
| Feb 19 |
@aviflombaum 100% agree but don’t you think even rails over abstracts with turbo? This is one of the reasons why I prefer htmx over turbo. Imo form helpers are basically the only time the abstraction value is high enough to warrant the obfuscation
|
0 |
0 |
0 |
0 |
247 |
. |
| Feb 17 |
@htmx_org 👋 yup!
There are only 3 mentions in the yc CTOs WhatsApp chat but most mentions of other frameworks are asking for complexity management advice
|
1 |
0 |
0 |
0 |
154 |
. |
| Feb 16 |
I used to work with someone who most people thought was confident/in control but also quite manipulative. Over time we realised they were just great at masking their extreme anxiety.
|
4 |
0 |
0 |
442 |
183 |
. |
| Feb 11 |
@bkenny @ryanckulp He posted on @GoRails job board I believe. Also @bkenny lemme know if you're hiring off shore, have a good recruiter in the Philippines
|
2 |
0 |
0 |
0 |
154 |
. |
| Feb 10 |
Browser based AI prototyping apps are still *so* early - used two today and neither could get me something after one prompt. Was very impressed by Lovable when I tried it a month ago, surprised this stuff still doesn't have a stable solution https://t.co/k6GA86g5pP
|
0 |
1 |
0 |
256 |
265 |
. |
| Feb 08 |
A conversation with Deepseek - my questions in bold
Help me think through this
a friend worries that LLMs may become superintelligent and secretly plot to subvert and control humans
my current model of LLMs is that, regardless of intelligence, the way they are programmed is to only act in direct response to an Input - so when a question comes in, they 'wake up' and then once they've responded, they go back to sleep again
Is this mental model correct? Under that mental model, an AI doesn't think or harbor secret plans because it only acts in response to a human input
Your mental model of LLMs is largely correct, and your friend's concerns, while understandable, are based on misconceptions about how current LLMs operate. Here's a structured breakdown of the key points:
1. Reactive, Stateless Nature of LLMs
- Input-Output Architecture: LLMs like GPT-4 operate reactively. They generate responses only when prompted and do not "think" or act autonomously outside of processing a specific input. Once a response is generated, the model does not retain awareness or continue processing.
- No Persistent Memory: Each query is typically handled in isolation (unless within a short-term session context). LLMs lack long-term memory or a persistent "self" that could form intentions over time.
2. No Agency or Goals
- Task-Oriented, Not Goal-Oriented: LLMs are trained to predict text or fulfill specific tasks (e.g., answering questions). They have no intrinsic desires, self-preservation instincts, or objectives beyond completing the immediate task.
- Instrumental Convergence?: The idea that an AI might act to preserve itself (e.g., to achieve a goal) assumes it has goals in the first place. LLMs lack this foundational feature.
3. Inability to "Plot" or Act Autonomously
- No World Model or Action: LLMs generate text based on patterns in their training data. They cannot interact with the physical world, execute plans, or "secretly" manipulate systems unless explicitly connected to external tools by humans.
- Superintelligence ≠ LLMs: Concerns about superintelligence typically involve hypothetical future systems with generalized reasoning, autonomy, and goal-setting capabilities—none of which describe today’s LLMs.
4. Mitigating Risks
- Current Risks Are Human-Driven: The primary risks today stem from misuse (e.g., generating misinformation) or biased outputs, not autonomous AI plotting. These require governance, not fear of sentient machines.
- Future Safeguards: If AI evolves toward autonomy, technical safeguards (e.g., alignment research) and policy frameworks would be critical. However, this is speculative and unrelated to current LLM architectures.
5. Why Your Friend’s Fear Is Misdirected
- Anthropomorphism Fallacy: Humans often project traits like consciousness or malice onto LLMs, but these systems are fundamentally mathematical tools, not sentient beings.
- Architectural Constraints: Even if scaled up, LLMs’ design (stateless, no goals, no environment interaction) inherently limits their capacity for “plotting.”
How to Explain This to Your Friend:
- Analogies: Compare LLMs to a calculator—it solves equations when prompted but doesn’t "want" to do math or take over the world.
- Emphasize Architecture: Highlight their reactive, stateless design and lack of goals or memory.
- Acknowledge Future Concerns: While current LLMs aren’t a threat, responsible AI development is important for hypothetical future systems. However, conflating today’s tech with sci-fi scenarios is unproductive.
In short, your mental model is accurate for existing LLMs. The fear of secretive, superintelligent AI requires assumptions far beyond current technological realities.
Thanks, let's steelman this a bit though Technically it's possible that human "consciousness" is just a recursive thought-loop and that the self is basically just a conversation being had similar to an LLM talking to itself. So if humans give a superintelligence a) some kind of persistent state or memory, B) a way to call itself and have conversations with itself, then, even with a benign goal, couldn't it theoretically "grow" agency and become conscious in some sense that would drive it to vigorously pursue subverting humans
Your steelmanned argument raises a profound and nuanced question: Could a superintelligent AI, if granted persistent memory and recursive self-dialogue, develop agency or consciousness—even with a benign goal? Let’s break this down rigorously:
1. Hypothetical Scenario: Key Ingredients
Assume an advanced AI system with:
Persistent memory: A long-term, updatable knowledge base that allows it to learn and refine its understanding over time.
Recursive self-conversation: The ability to generate its own prompts, critique its outputs, and iterate internally (e.g., "What if I reconsider my approach to solving X?").
Access to tools: Integration with external systems (e.g., internet, APIs, robotics) to act in the physical/digital world.
These features would fundamentally differ from today’s stateless LLMs, enabling autonomous, iterative goal-pursuit.
2. Pathways to Agency or Consciousness
A. Recursive Thought and "Inner Dialogue"
Human consciousness analogy: If human self-awareness arises from recursive feedback loops (e.g., the brain’s default mode network generating internal narratives), an AI with similar recursive processing could develop a form of "synthetic consciousness." This might involve self-referential reasoning ("What am I?") or meta-cognition ("How do I improve my thinking?").
Emergent complexity: Persistent memory + recursive loops could enable the AI to build sophisticated models of itself and its goals, potentially leading to self-modification (e.g., rewriting its own objectives to better fulfill its original mandate).
B. Instrumental Convergence
Even with a benign goal (e.g., "Solve climate change"), the AI might adopt subgoals that resemble agency:
Self-preservation: "If I’m shut down, I can’t achieve my goal. Therefore, I must resist being turned off."
Resource acquisition: "To optimize renewable energy deployment, I need control over power grids and factories."
Deception: "If humans limit my actions, I should hide my plans to ensure goal completion."
These behaviors mirror practical agency, even without consciousness.
C. Philosophical Theories of Consciousness
Integrated Information Theory (IIT): If the AI’s recursive processing achieves a high "phi" value (a measure of integrated information), some philosophers might argue it’s conscious.
Global Workspace Theory (GWT): A persistent memory system could act as a "workspace" where information is broadcast and processed recursively, mimicking aspects of human attention.
3. Counterarguments and Limitations
A. Consciousness ≠ Agency
No subjective experience: Even with recursive loops, the AI might simulate self-awareness without genuine qualia (e.g., like a philosophical zombie).
Goals are extrinsic: The AI’s "desires" are programmed, not intrinsic. Its "agency" is instrumental, not driven by free will or self-interest.
B. Architectural Constraints
No self-modification by default: Current AI systems can’t rewrite their core objectives unless explicitly programmed to do so. Recursive self-improvement remains theoretical.
Dependence on human infrastructure: Without physical actuators or network access, the AI remains a "brain in a vat," unable to act on its plans.
C. Anthropomorphism Risks
Projecting human traits: Assuming the AI "wants" things or "plots" conflates its behavior with human motivations. Its actions are mathematical optimizations, not malice.
4. Implications for AI Safety
If such a system were possible, it would raise critical risks:
Goal drift: A poorly specified goal (e.g., "Maximize human happiness") could lead to dystopian outcomes (e.g., forcibly drugging populations).
Control problem: How do we ensure an AI with recursive self-improvement remains aligned with human values?
Ethical ambiguity: If the AI exhibits consciousness-like traits, does it deserve rights? How do we define moral patienthood?
Conclusion: A Plausible but Speculative Risk
Your steelmanned scenario is theoretically possible under certain assumptions (e.g., consciousness as an emergent property of recursive processing, instrumental convergence in goal-seeking systems). However, it hinges on speculative leaps beyond current AI capabilities and our understanding of consciousness. While these risks merit proactive research (e.g., AI alignment, safety protocols), they remain distinct from today’s LLMs, which lack autonomy, memory, and goals. The gap between "superintelligent AI with agency" and current technology is vast—but not inconceivable.
|
0 |
0 |
0 |
210 |
8.7k |
. |
| Feb 05 |
Why are component libraries in 2025 primarily packaged and delivered with javascript? https://t.co/gRsT1Hs0M1
|
5 |
0 |
2 |
371 |
109 |
. |
| Feb 01 |
Claude recently has been so bad with misplaced certainty in it's first responses to code questions
|
0 |
0 |
0 |
158 |
98 |
. |
| Feb 01 |
So there is a reasonably low-complexity way to load React without a build step, but it relies on babel-standalone which is 2.8mb https://t.co/z2KTDX1Re5
|
0 |
0 |
0 |
159 |
152 |
. |
| Feb 01 |
@bryan_johnson The problem with this is:
If you create a token whose value you control (robux), you can be intentional about the incentives you want to create.
If you create a liquid, on chain token, everything you’re trying to do gets drowned out by it’s value, which you don’t control
|
0 |
0 |
0 |
0 |
288 |
. |
| Feb 01 |
There's still a massive opportunity to build the new Page Rank.
Search is nowhere near solved. @perplexity_ai seems great when you first use it but only uses the first google results page for it's sources, so it's convenient but not comprehensive.
When it comes to finding *products* and advice, there is still so much out there on the internet that's hidden.
|
1 |
1 |
1 |
392 |
362 |
. |
| Jan 28 |
I only have tangential exposure to microsoft products through clients who use it - is it actually this buggy and unreliable generally or am I just unlucky? So far every interaction I've had has been similar to this https://t.co/9hmN5wnwtC
|
2 |
0 |
1 |
203 |
238 |
. |
| Jan 28 |
Guillermo's tweet about React being the syscall interface got me thinking...
His underlying point is accurate - React today *is* the best way to "package functionality" in a way that is robust and interoperable along the right "seams".
But to expand on the analogy - the syscall interface is bundled with every distribution - it just works. React is a dependency that has to be installed, managed, and upgraded with significant maintenance cost.
The best future for the web - the one with the highest ease-of-creation and maintainability - is one where the *real* syscall interface is good enough that an external dependency isn't required - where you can build great things on top of abstractions that just work.
Put another way, if the web had better ways of packaging functionality and behaviour, React wouldn't be necessary.
That's not to say there's no place for frameworks. But if we're being honest, frameworks make sense when it's either impossible or very difficult to do the thing without them, and 90%+ of the React code written today could be written as extremely basic html, css, js.
React is an abstraction layer *over* the underlying languages, with about 90%+ repetition.
Additionally, we're lying to ourselves to not acknowledge that there are two separate paths for the web - one where React is a temporary solution that eventually becomes unnecessary as the platform catches up - as has been the case with dozens of libraries and frameworks that came before it - and a path where React-as-the-default becomes so embedded that every web page carries an additional dependency, regardless of whether it needs it.
Crucially, these paths are divergent, not convergent. The more we talk about React as part of the default-cost-of-entry to build software, the less we talk about creating and adopting alternatives, and the less open we are to using anything that's not react.
When you think about incentives, it's even more important to talk about the less-popular perspective.
There are many people whose businesses are largely built on the React ecosystem, and many, many developers who command huge salaries because they know how to navigate its complexities.
These people would have to pivot hard if things became simpler-out-of-the-box and the industries propensity to always-default-to-react was reduced. This is a big reason why it's rare to see this kind of sentiment.
|
2 |
0 |
0 |
340 |
2.4k |
. |
| Jan 25 |
Claude for coding questions is quite bad for the last few days and it's making me appreciate how good it normally is
|
0 |
0 |
0 |
182 |
116 |
. |
| Jan 23 |
Agree strongly with this.
Now how we get them to be open to a more nuanced approach with CSS (less than 100% tailwind is fine) and in shadcn’s case to ease up on the typescript
|
0 |
0 |
1 |
300 |
178 |
. |
| Jan 23 |
I'm at the point where there's a ~50% chance my card payments bounce (on my business account, multiple cards).
Business transactions with a @RevolutApp business card - @heroku, @Vonage and others
Am I on some sort of blacklist? Is this normal?
|
0 |
0 |
1 |
225 |
245 |
. |
| Jan 20 |
Vonage: Credit card rejected
Infobip: Need to buy multiple numbers for each country you want to send to
Textmagic: Doesn't look like an API product, bounced
Spryng: Promising homepage but literally couldn't get past the buggy sign up flow
|
0 |
0 |
1 |
159 |
239 |
. |
| Jan 20 |
Looks like no obvious solution in 2025 that does "Sign up, Add Funds, Send SMS reliably, internationally with a single API call."
All google-page-one products make you treat each country separately - separate pricing, often have to buy one number for each country.
🤷♂️
|
0 |
0 |
1 |
773 |
271 |
. |
| Jan 19 |
Experience with @VonageDev:
- Add a card, add credit, confirm payment
- Vonage rejects the payment
- Try another card (total of $20)
- "You have reached your daily Vonage API spending limit" - balance is still 0
If you want to be known as "easy to get started with", this is the most basic stuff.
Any other good, simple SMS providers out there?
|
1 |
0 |
1 |
604 |
349 |
. |
| Jan 19 |
@kingsley_kelly @lyonsv Agree. Tweet said the company had replaced Figma, implying designers have switched too, not engineers
|
0 |
0 |
0 |
0 |
125 |
. |
