| Jun 14 |
The more I read anecdotes on this, the less I trust the US and people like Sacks on AI. No one has defined what “jailbreak” means in this context, and as the AI czar it’s pretty important to understand and be able to communicate (with other AI companies and the public)
|
1 |
0 |
0 |
172 |
269 |
. |
| Jun 12 |
This 100% - I am *baffled* by all the "I just leave my agent running for hours/overnight and come back to finished software" - there's no way that can create good software, is there?
|
3 |
0 |
1 |
169 |
182 |
. |
| Jun 12 |
I think a lot of people - myself included - use twitter as their "cynicism release valve", and as a result come across much more grumpy than they are IRL. Watching @thekitze go from moderately cynical but still entertaining - to as of recent - unabashed criticism and negativity - is a cautionary tale in what happens when you lean into it too much.
|
0 |
0 |
0 |
91 |
349 |
. |
| Jun 10 |
Sinertia
Sinatra on the backend, activerecord for ORM, preact on the frontend, inertia in the middle - is my go-to stack for AI assisted coding
- Super fast to generate
- Close to zero ambiguity
- Power of a real backend language
- Extremely easy to code review
- Rarely breaks
- Apps feel extremely snappy
💯
|
2 |
0 |
1 |
174 |
313 |
. |
| Jun 10 |
Doing a lot of hardening of several largely-vibecoded features and the thing that keeps biting me is the *fallbacks* - instead of just raising an error there'll be some weird conditional that tries to get data from somewhere else - not sure how to prompt this away, seems to be something that both claude & openai do
|
0 |
0 |
1 |
90 |
316 |
. |
| Jun 06 |
Still using all three - Claude Code, Cursor, and Codex - on a day to day basis
- Codex 5.3 spark for iterating on UI & simple backend
- Claude Code (in the Codex app terminal) for when things get architectural
- Cursor for when I need to drop down and explore things myself.
|
0 |
0 |
1 |
208 |
279 |
. |
| Jun 01 |
The *moment* we broke our single-ruby-codebase out into a control plane and a runner, claude code (Opus 4.7) became close to useless - not able to follow a chain of connections when diagnosing and fixing - modifying files-in-place, not syncing or restarting - basically no theory of how changes flow through the system whatsoever
|
1 |
0 |
1 |
176 |
329 |
. |
| Jun 01 |
@pewdiepie enters the chat https://t.co/3Vcw2yFuBV
|
0 |
0 |
0 |
82 |
50 |
. |
| May 28 |
How you know it's possible for the majority to be wrong
😝 https://t.co/6bz5y5Ers2
|
1 |
0 |
0 |
140 |
82 |
. |
| May 27 |
Absolute State of the Art speed wise for LLM assisted software dev: A single file sinatra backend driven by codex-5.3-spark (on the $200 month plan) at 1000 tokens/second. I'm not sure it can get much better than this. Thanks @sama @thsottiaux https://t.co/fstwviBuUN
|
3 |
0 |
1 |
377 |
267 |
. |
| May 23 |
This shit is exhausting. Switched to Codex because 5.3 spark & @thekitze said the remote ssh stuff was better than cursor, but they have the same bugs as the others - in this case upgraded my account but somehow the conversation stays locked to "you've hit your usage limit" https://t.co/YLRTNKVWe0
|
2 |
0 |
0 |
261 |
302 |
. |
| May 23 |
Claude Code's AskUserQuestion tool in my experience 9/10 times is less useful than just asking the questions conversationally
|
3 |
0 |
1 |
293 |
125 |
. |
| May 22 |
Someone... didn't think through the UX here https://t.co/kv01h9D92w
|
1 |
0 |
2 |
169 |
67 |
. |
| May 19 |
4 months from this tweet to Viktor going mega viral - they are now the Lovable of the category. In the mean time we built an almost identical product (Gigabrain) but added far more web features on top, which was probably unnecessary in hindsight.
Anyway, now the Viktor folks are out of the gates it’s time to get after them - in the same way there were multiple 9 figure revenue vibe coding companies, there will be multiple “ai co worker” companies too.
I expect to be tweeting less about random dev stuff and a lot more about Gigabrain going forward.
|
3 |
0 |
0 |
469 |
556 |
. |
| May 19 |
This feels like the first milestone of the next step-change - open source models that are close to opus-level but 10x as fast. Now we just need bigger context window, and more providers than Cerebras. Big
|
5 |
0 |
1 |
432 |
204 |
. |
| May 14 |
Low status practices can be harmful if you have people with limited experience finding the edges of systems and understanding the harms that best practices protect from. But an extremely powerful weapon is the experienced practitioner who's ok with (thoughtful) low status ideas
|
2 |
0 |
0 |
228 |
278 |
. |
| May 13 |
...
https://t.co/eUOGZhMnhU
|
0 |
0 |
0 |
98 |
29 |
. |
| May 12 |
https://t.co/tH3HtefvfZ
|
1 |
0 |
0 |
84 |
23 |
. |
| May 11 |
This is the difference between a vibe-coded-in-a-weekend product and spending months dogfooding & ironing out every little kink https://t.co/l2pf6sx4Dx
|
3 |
0 |
0 |
206 |
155 |
. |
| May 10 |
Spent the day using codex-5.3-spark (had to dust off cursor) - this feels like a genuine step change.
My guess is a lot of code bases need a wider context window or better reasoning, but our sinatra backend, no-react frontend it is *both* insanely fast and extremely capable.
|
3 |
0 |
0 |
194 |
277 |
. |
| May 06 |
This 1000x. If you’ve been actually doing the work through chat, and extracting snippets or summaries as json blobs from every exchange, you’ll know that this is where the magic happens
|
1 |
0 |
1 |
318 |
185 |
. |
| May 06 |
https://t.co/scvAvL162a
|
0 |
0 |
0 |
120 |
23 |
. |
| May 06 |
https://t.co/scCGAgKt4D
|
1 |
0 |
1 |
146 |
23 |
. |
| May 06 |
Every day I see two to three tweets in my feed of things we're solving with Gigabrain (fka Operator) - going to collect them all here so that we can make some content on them when we (finally) go public
|
3 |
0 |
2 |
194 |
202 |
. |
| May 06 |
December 2025 will go down as the single month where hundreds, if not thousands, of product & AI-minded developers got Claude Code pilled and began building their own prototypes based on the below realization, many of which are now starting to pop up on our twitter feeds.
|
2 |
0 |
0 |
274 |
276 |
. |
| May 06 |
Also not to toot my own horn, but if you've been following me since Jan you'll have known this for quite some time.
https://t.co/cnSOQNiNaz
|
0 |
0 |
0 |
53 |
141 |
. |
| May 06 |
AI Brain
AI Workspace
AI Operating System
Are the next mega-category of AI product
https://t.co/rYZNglWILt
|
0 |
0 |
1 |
123 |
108 |
. |
| May 06 |
@blakeandersonw @HilaShmuel @dotta Welcome @jacob_posel
https://t.co/jbOlF7p3vv
|
1 |
0 |
1 |
116 |
81 |
. |
| May 05 |
There was a day when Twitter content looked like this and I miss it - there’s something so satisfying about finding non-obvious observations and compressing them as far as they’ll go without losing the insight. I get 2 or 3 tweets a day in my feed like this nowadays and they’re all from Justin
|
3 |
0 |
0 |
230 |
294 |
. |
| May 04 |
Of course @zapier was also working on an AI workspace product
https://t.co/zyql9ryAJo
|
0 |
0 |
0 |
86 |
87 |
. |
| May 02 |
I feel personally targeted by this tweet
|
1 |
0 |
1 |
175 |
40 |
. |
| May 01 |
@annimaniac you may be interested in this :)
https://t.co/eCqYHgwMPR
|
1 |
0 |
0 |
113 |
70 |
. |
| May 01 |
There's a new article like this every day on X, but this is the best one I've seen, and the one that most aligns with my experience
|
4 |
0 |
1 |
425 |
131 |
. |
| Apr 27 |
DevX maybe?
|
0 |
0 |
0 |
182 |
11 |
. |
| Apr 27 |
This is takeaway #1 and #2, but takeaway #3 is that the frothing at the mouth over how intelligent the models are, and what that means for the future of software engineering is in harsh contrast against what everyone is experiencing in practice. It’s like having an F1 driver as your uber driver who will occasionally glitch out and drive you off a cliff.
|
1 |
0 |
0 |
177 |
355 |
. |
| Apr 25 |
When I chat to my adhd friends - we all have super weird/esoteric strategies for dealing with it/maximising output, but what's fascinating is there's basically zero overlap & we all look at each other funny when we explain how we do stuff
|
2 |
0 |
0 |
187 |
243 |
. |
| Apr 14 |
I really wish otherwise super smart people would stop contributing to extremely silly narratives (running agents for a long time = a good thing) that other people and the media then also buy into. Just because something sounds cool doesn’t mean it makes sense. This is dumb
|
2 |
0 |
1 |
393 |
273 |
. |
| Apr 09 |
People building new models should stop benchmark-maxxing and just focus on "how does this feel when it's hooked up to Code or Cowork", because that's by far the most important way that these models are going to be wielded.
|
3 |
0 |
0 |
265 |
222 |
. |
| Apr 06 |
😱 Bitcoin price is daaangerously close to $69,420 https://t.co/SLosNckXFz
|
1 |
0 |
0 |
77 |
73 |
. |
| Apr 06 |
Thoughts on Openclaw
- It's created a new product category in consumers minds - AI chat that has great context and can *do* stuff (as opposed to just chat) = 👍
- At the same time, messaging that OpenClaw is now the De Facto way to get AI to do stuff, products being touted as "built for Openclaw" vs "built for agents" = 👎
|
0 |
0 |
1 |
235 |
322 |
. |
| Apr 03 |
@MediaKing eli5?
|
1 |
0 |
0 |
118 |
16 |
. |
| Apr 03 |
Matt's the expert here & i know very little, but I really struggle to understand this equation. Stripe doesn't just *get* the Indiehackers audience, nor does Hubspot with The Hustle - they can run ads for free, but editorial is still the same - where does this materialize?
|
1 |
0 |
1 |
351 |
277 |
. |
| Apr 03 |
🤣 https://t.co/6Zpeh5ycaw
|
1 |
0 |
0 |
170 |
25 |
. |
| Apr 03 |
It took one day 😆
https://t.co/TxbxijFCZA
|
0 |
0 |
0 |
105 |
43 |
. |
| Apr 02 |
Haven't seen anyone talk about this - you can now pay to bypass ads on Instagram? https://t.co/8Zq2K8gTwI
|
1 |
0 |
0 |
212 |
105 |
. |
| Apr 01 |
We now have 2 models that *feel* on par with Opus inside Claude Code. The glaring gap is still context window - compaction is still super regular and annoying. Once someone brings out a model that has same performance with the 1m context window, we'll have a realistic replacement.
|
2 |
0 |
1 |
336 |
281 |
. |
| Mar 31 |
This is such a great test for exploring "the end of labour" with AI - building something great is the product of a lot of seemingly small things that add up - do people really think that all the items on this list are *fully replaceable* by AI, even in the next 10 years?
|
0 |
0 |
0 |
227 |
271 |
. |
| Mar 26 |
Seeing and taking this has completely destroyed the credibility for me of multiple people I held in very high esteem, who have either been gesturing strongly or stating explicitly that LLMs will get us to AGI - it makes the whole thing seem just completely farcical
|
1 |
0 |
0 |
196 |
265 |
. |
| Mar 26 |
I always thought having *multiple* AI bot identities didn't make sense - if intelligence-capacity isn't the constraint, why not just have one "God" bot that can handle everything, but I realized it's actually the best mechanism for setting context. E.g. I know that this identity has these credentials and access to these instructions.
|
1 |
0 |
0 |
114 |
335 |
. |
| Mar 23 |
https://t.co/JgBHhPpE28
|
0 |
0 |
0 |
96 |
23 |
. |
| Mar 20 |
Another AI workspace enters the chat
https://t.co/Kp4Avvi6Z8
|
0 |
0 |
0 |
277 |
61 |
. |
| Mar 16 |
I keep seeing this caricature of men and I have to say, I literally don't know anyone who acts like this? Reminds me of the way the manosphere guys caricature women.
Is this a US thing? Am I in a bubble? Are other people in a bubble?
|
0 |
0 |
0 |
311 |
234 |
. |
| Mar 15 |
Something I can see happening in the near future
- Agents cause an increasing number of security issues & also begin to hammer APIs
- Products begin to lock down their APIs, require 2fa, or charge extra for API/MCP
- People end up just building their own replacements
|
4 |
0 |
1 |
246 |
273 |
. |
| Mar 12 |
The dumb thing about this is that the MCP spec could have allowed "rest API with Oauth" as a method, but they mandate the ridiculous every-tool-is-a-local-server approach, and now people are rightly seeing it's too burdensome and skipping it completely
|
0 |
0 |
0 |
244 |
253 |
. |
| Mar 11 |
Hallucinations starting to go back up again - wasn't this meant to be solved? https://t.co/pqDDsmM7A2
|
1 |
0 |
1 |
111 |
101 |
. |
| Mar 08 |
At some point companies will compose 100% of their execution from modular processes that will be gradually converted from human-only, to human-in-the-loop, to agent only, and you'll be able to visualize, manage and monitor your company just like an n8n workflow https://t.co/zcsBPqumxy
|
1 |
0 |
1 |
366 |
285 |
. |
| Mar 08 |
Would like to keep a thread of interesting people working on ZHCs (Zero Human Companies) or similar.
So far I have
@Shpigford
@nateliason
@johnrushx
@Bencera
@tonyennis (yours truly)
Anyone else to keep an eye on?
|
2 |
0 |
1 |
347 |
220 |
. |
| Mar 08 |
I am simultaneously...
- Building the most insane autonomous AI stuff & giddy about what is possible.
- Having extreme doubts about the "death of white collar work" and quote-unquote "intelligence" of the current models
I would guess same for a lot of people
|
1 |
0 |
2 |
290 |
264 |
. |
| Mar 06 |
https://t.co/QbEPxXLNee
|
0 |
0 |
0 |
79 |
23 |
. |
| Mar 06 |
Direct quote from a senior consultant my client brought in, on a call this week: “You can’t just keep cramming things on a server, it’s just not how things scale”. It was a 2cpu 4gb ram server for internal tools that was at about 40% capacity on most metrics. The client has a few dozen customers and 50 projects.
🤷♂️
|
3 |
0 |
0 |
319 |
319 |
. |
| Mar 06 |
Enjoy Derek's takes but...
- "Boastful" certainly makes it easier & that mightn't be a bad thing. This culture is partly why the US is so prosperous
- "Sham Products" - you could write an entire book on the delicate moral philosophy of what constitutes valuable. The underlying questions are 1. "What percentage of people who buy actually get value", and 2. At what ratio do you consider something "a sham". But many things have the characteristic that people who don't get value from them outnumber people who do get value - often it's the entire business model - gyms, insurance, social security.
- "You can easily be a millionaire" - Ask anyone who owns a business that has employees, customers, suppliers etc & sees the effort required to hit seven figures in revenue, the difference between topline and the amount you actually take home, and how much effort is and stress is required, & you'll be very hard pressed to find anyone who would describe it as easy.
Of course grifters exist, but this level of puritanical cynicism affects people's worldviews in a way that prevents a lot of people who have something of genuine value to offer (e.g Derek himself), from doing so, and leads to black-and-white, spiteful thinking that penalizes any form of self promotion or monetization that are by far net positive.
|
1 |
0 |
0 |
177 |
1.3k |
. |
| Mar 03 |
AI (Claude) is still not good at well reasoned, well structured argumentative documents, particularly high stakes ones.
But speaking to AI and getting a first draft that you can completely rewrite is still a helpful exercise.
|
0 |
1 |
0 |
120 |
227 |
. |
| Mar 03 |
https://t.co/1xHkwYBcrE
|
0 |
0 |
1 |
171 |
23 |
. |
| Mar 02 |
This is awesome. If you want to take it to the next level and have both modern approaches and beautiful styling, https://t.co/lWxjnoglNd
😁
|
0 |
0 |
0 |
165 |
139 |
. |
| Mar 01 |
https://t.co/qe09e6twV9
|
0 |
0 |
1 |
80 |
23 |
. |
| Feb 21 |
I found an in between here
Started with db backed Ui that had lots of features but was fundamentally point and click
Ended up with a web based file browser - a thin layer on top of the filesystem that gives me extra stuff I don’t get with Finder. Now I use chat to do the acting but I can explore and inspect the outputs or intermediate steps where necessary
|
0 |
0 |
0 |
325 |
360 |
. |
| Feb 20 |
Is anyone using Notion agentically to store content? Seems incredibly inefficient with claude code https://t.co/j1orLmSk8D
|
1 |
0 |
0 |
218 |
122 |
. |
| Feb 19 |
Something the "Software is dead" commentary misses is the sheer premium people will pay to not have to manage or worry.
It's been possible for years to get software built cheaply on Fiver or Upwork, but it takes *a lot* of back and forth, the responsibility burden for quality, stability etc is borne by the manager, not the doer, peace of mind is lower, and likelihood of success is lower. So people pay an agency 6 or 7 figures to build the same proof of concept.
AI obviously changes the equation, but I'm not sure it does so substantially enough to remove those considerations entirely.
|
2 |
0 |
1 |
175 |
593 |
. |
| Feb 15 |
This was probably the most impressive thing Claude Code has done for me so far, there's no way I would have figured this out without spending days on it. https://t.co/fOXlhOP3Ja
|
3 |
0 |
2 |
316 |
177 |
. |
| Feb 09 |
https://t.co/3CNrELJc87
|
0 |
0 |
0 |
90 |
23 |
. |
| Feb 08 |
Ooooh https://t.co/VpAIZbFNbx
|
0 |
0 |
0 |
93 |
29 |
. |
| Feb 06 |
https://t.co/tD27v9Zq5Y
|
0 |
0 |
1 |
218 |
23 |
. |
| Feb 06 |
Banger tweet
|
2 |
0 |
0 |
141 |
12 |
. |
| Feb 04 |
More on this. Using Pi as the harness...
- Kimi k2 on Groq is fast but a bit dumb, and dangerous
- GLM on Cerebras just has non-stop API rate limiting issues
But both give you a taste of Claude Code at near instantaneous speed & it feels crazy
Looks like post acquisition @GroqInc has no plans to add Kimi 2.5
Are there other super fast inference providers?
|
0 |
0 |
2 |
431 |
363 |
. |
| Feb 04 |
Aaaaand I lost a half day of work 🙃
Kimi k2 is fast but if you get to used to Opus/CC you forget these models can take extremely destructive actions on a whim (like git reset --hard HEAD for example)
|
0 |
0 |
0 |
154 |
200 |
. |
| Feb 03 |
Switching from Claude Code to Pi wired up to Kimi K2 on Groq is like going from listening to podcasts on 1x to 2.5x.
|
0 |
0 |
0 |
399 |
116 |
. |
| Feb 03 |
Old enough to remember when @openclaw was called Clawdis
|
0 |
0 |
0 |
102 |
56 |
. |
| Feb 03 |
This is the exact inverse of my takeaway from this. There was insane levels of hype all the way through from 2020, and the sane prediction at the time was actually “This will take 5 or 6 years before it changes anything meaningfully” . In fact if you’d said that, a majority of people would have laughed at you for underestimating and mocked you for not seeing the future.
|
0 |
0 |
0 |
184 |
372 |
. |
| Feb 03 |
Second js-based app that's crashing on me this week. @convex what's the deal? My browser literally can't handle your SPA https://t.co/cNCSuQbhEd
|
0 |
0 |
0 |
127 |
144 |
. |
| Feb 03 |
This. I spent a few hours yesterday trying to get glm on cerebras or kimi on groq wired up and I can’t help but feel there are incentives at play here to make this very hard and normalise the slow, closed source expensive models
|
0 |
0 |
0 |
194 |
228 |
. |
| Feb 03 |
Stick it on a vm, add passenger to your rails apps, connect cursor or vscode directly to it and you now have “live development”. We’ve been doing this for years & it’s awesome
|
0 |
0 |
0 |
171 |
179 |
. |
| Feb 02 |
https://t.co/RJ7euWCE5F
|
0 |
0 |
1 |
102 |
23 |
. |
| Feb 02 |
Really bums me out how abysmally bad the accounting/tax industry is when it comes to customer experience
I've used 10+ firms across jurisdictions & have yet to meet anyone who has more than 2/10 focus on assisting the client - helping you understand, deal with bureaucracy, find the best solution.
It's basically just "Here's a list of shit you have to do, scattered over 10 email threads". "Oh you want to do *that*, you'll have to figure it out yourself".
Once you're outside the happy path (e.g. any kind of international setup) you're completely on your own.
|
0 |
0 |
0 |
117 |
566 |
. |
| Feb 02 |
Got a cold email from a sushi restaurant.
Not sure if this is a one-off (insanely resourceful marketing team) or if AI is making this much easier and we can expect a much wider range of cold emails in future. https://t.co/mvltEqT1ol
|
0 |
0 |
0 |
129 |
234 |
. |
| Feb 02 |
https://t.co/9cGXAyNbkx
|
0 |
0 |
1 |
95 |
23 |
. |
| Jan 31 |
Didn't expect when I tweeted this that it would only take a week for the first breakout consumer example (@openclaw) to hit the mainstream.
|
0 |
0 |
1 |
266 |
139 |
. |
| Jan 30 |
Took a week for me to realize how much of a game changer @flydotio new product is
Haven't used it yet because I'm still building my agent in solo mode, but for AI apps that need their own VM this is going to remove so much pain.
https://t.co/KTxq8z8myK
|
1 |
0 |
0 |
224 |
256 |
. |
| Jan 29 |
I have a hunch that a huge contributor to nosediving fertility rates is a shift in what's generally considered as the bar for being a "standard good" parent. People who have kids non-stop-humblebrag on socials about what they do for them and how they center their life around them. People who don't have kids see where the bar is, and go "I don't know if I can meet that".
From what I gather for most of history people didn't overthink having kids. Now it seems like it has to be a "hell yes or no" situation.
|
1 |
0 |
1 |
240 |
510 |
. |
| Jan 28 |
What’s the thinking here? That the government should prevent employers from hiring overseas?
|
0 |
0 |
0 |
190 |
92 |
. |
| Jan 23 |
... https://t.co/QVT6g2BSyU
|
0 |
0 |
0 |
121 |
27 |
. |
| Jan 20 |
Looks like @clickup acquisition ruined @codegen 😢
|
1 |
0 |
0 |
164 |
49 |
. |
| Jan 19 |
*bare 🤦♂️
|
0 |
0 |
0 |
93 |
10 |
. |
| Jan 19 |
Coding is maybe the perfect domain to lay bear the difference between having a huge amount of information and the ability to construct semantically correct language, and actually being smart. It's still in the harness, not the models. https://t.co/uImyFpKSuI
|
0 |
0 |
2 |
244 |
258 |
. |
| Jan 17 |
https://t.co/fIf1JZHKMa
|
0 |
0 |
1 |
145 |
23 |
. |
| Jan 17 |
Good example of what I mean https://t.co/dq1PMYjwPd
|
0 |
0 |
1 |
143 |
51 |
. |
| Jan 15 |
Consumer focused vibe coding app business models are the digital equivalent of a gym chain
90% of people will join aspirationally but never follow through enough to see any real value - because the hard part is getting people to pay you for what you built.
|
0 |
0 |
0 |
101 |
258 |
. |
| Jan 15 |
The only reason people have to do this now is because the agents are still slow. A lot of the fanfare about multiple parallel agents will look comical in hindsight
|
0 |
0 |
1 |
133 |
163 |
. |
| Jan 13 |
Best product I've seen doing this so far is @zocomputer - very cool stuff.
|
1 |
0 |
0 |
147 |
74 |
. |
| Jan 12 |
... https://t.co/4St9baG6fg
|
0 |
0 |
0 |
66 |
27 |
. |
| Jan 11 |
This is one of the dumbest things I’ve seen a product leader say in a long time and makes me quite bearish on Twitter. “The product” for 95% of people is the tweets they see. Why would you lead product if you don’t own or have control over 90% of it
|
1 |
0 |
0 |
220 |
249 |
. |
