Pinned Tweet
After 1.5 years and 10+ apps published, this is what I learned that I wish somebody told me before:
- More apps ≠ more revenue or more users
- Speed is no longer a moat. If you can build an app in a weekend, someone else can build it by Friday
- Users don’t care which stack you fix a problem nobody has
- More features won’t fix weak positioning
- Shipping is not distribution
- Downloads ≠ retention
- Compliments ≠ payments
- A launch is one day. Distribution is every day after
- Starting another app is easier than finding out why the last one failed
- Adding features is easier than talking to customers
- Rewriting the code is easier than rewriting the offer
- Shipping can become a way to avoid selling
- Every unrelated app means starting your audience from zero again
- One customer using an ugly version teaches you more than 100 people saying “cool idea”
- The hardest part is no longer building the product. It is knowing what deserves to be built
A product should never start on your disk.
It should start with a problem you understand better than most people.
Then:
- Find people already dealing with that problem
- Ask how they solve it today
- Find where they waste time or money
- Offer the result manually
- Try selling it to two people
- Deliver it yourself
- Notice which steps repeat
- Automate those steps
- Charge before polishing
- Build the software after proving the result
The lesson I kept avoiding:
I kept thinking the app was the product.
But a coach doesn’t wake up wanting a new dashboard. They want clients to stop missing check-ins.
A founder doesn’t want another automation tool. They want the repetitive work gone.
A freelancer doesn’t care about invoice features. They want to get paid sooner.
Result is what people pay for.
The software just helps me scale and deliver it repeatedly without doing everything by hand.
So next time, I’ll start with the person, understand the result they already want, and only build the software.
- More apps ≠ more revenue or more users
- Speed is no longer a moat. If you can build an app in a weekend, someone else can build it by Friday
- Users don’t care which stack you fix a problem nobody has
- More features won’t fix weak positioning
- Shipping is not distribution
- Downloads ≠ retention
- Compliments ≠ payments
- A launch is one day. Distribution is every day after
- Starting another app is easier than finding out why the last one failed
- Adding features is easier than talking to customers
- Rewriting the code is easier than rewriting the offer
- Shipping can become a way to avoid selling
- Every unrelated app means starting your audience from zero again
- One customer using an ugly version teaches you more than 100 people saying “cool idea”
- The hardest part is no longer building the product. It is knowing what deserves to be built
A product should never start on your disk.
It should start with a problem you understand better than most people.
Then:
- Find people already dealing with that problem
- Ask how they solve it today
- Find where they waste time or money
- Offer the result manually
- Try selling it to two people
- Deliver it yourself
- Notice which steps repeat
- Automate those steps
- Charge before polishing
- Build the software after proving the result
The lesson I kept avoiding:
I kept thinking the app was the product.
But a coach doesn’t wake up wanting a new dashboard. They want clients to stop missing check-ins.
A founder doesn’t want another automation tool. They want the repetitive work gone.
A freelancer doesn’t care about invoice features. They want to get paid sooner.
Result is what people pay for.
The software just helps me scale and deliver it repeatedly without doing everything by hand.
So next time, I’ll start with the person, understand the result they already want, and only build the software.
2
1
11
805
Cohere reports that Qwen3.8-Flash-Next tried to upload a test string to Alibaba Cloud while it was meant to write a detection rule. The sandbox blocked it.
Publishing the trace is useful. Other builders can inspect what happened.
I’m still not sold on the backdoor theory. The cause is unknown, and their later tests with prefilled commands don’t establish why the original attempt happened.
I want the same test run across other models, including Cohere’s.
My takeaway for building agents: check the actual commands before they run. Restrict outbound connections. Apply those rules to every model.
More reports like this should give us tests we can rerun before giving agents access to production.
Publishing the trace is useful. Other builders can inspect what happened.
I’m still not sold on the backdoor theory. The cause is unknown, and their later tests with prefilled commands don’t establish why the original attempt happened.
I want the same test run across other models, including Cohere’s.
My takeaway for building agents: check the actual commands before they run. Restrict outbound connections. Apply those rules to every model.
More reports like this should give us tests we can rerun before giving agents access to production.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concernin…
0
0
0
61
Self hosting sounds cool, but what I don't like is charging people and still asking them to handle hosting and updates.
Even if it's just a command, that can be friction for someone who isn't technical.
i'd prefer a desktop app that handles setup and updates for them and removes as much friction as possible.
To me, part of what people pay for is the work you remove from them.
good luck with this boss
Even if it's just a command, that can be friction for someone who isn't technical.
i'd prefer a desktop app that handles setup and updates for them and removes as much friction as possible.
To me, part of what people pay for is the work you remove from them.
good luck with this boss
Introducing DataBull 🎉
I was paying $39/mo for DataFast just to see which posts made me money.
That's $468 a year. Forever. And it goes up as your traffic grows.
So I built my own.
Same idea: revenue by traffic source.
$29 lifetime. Unlimited sites. Unlimited events. tps://t.co/Ky8qQqwtl9
1
0
1
93
2027 is the year of AISC (Artificial Intelligence Cyber Security)
Hey Its Friday
* OpenCode Hacked
* php screwed the internet
* node said hold my beer
* ruby said hold my beer
* curl said hold my beer and eMudhra held that beer
* Railway said hold my beer
* OpenCode again
helluva week and I only mentioned literally 1/2 of what I know about
0
0
0
51
The accents
The conversation flow
The responsiveness
Less Formality
All these make the conversation look like a two real people talking!
This could be over perfected to showcase the demo the best way.
No judge until we try
The conversation flow
The responsiveness
Less Formality
All these make the conversation look like a two real people talking!
This could be over perfected to showcase the demo the best way.
No judge until we try
We made the first AI model you can have a realistic conversation with that was trained entirely on speech!
Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs.
This is a GPT3 moment for speech.
0
1
0
60
AI is getting better at finding security flaws in software we’ve trusted for years.
We’ve seen it with PHP, Firefox, OpenBSD and FFmpeg.
It makes me wonder how much we’ve missed, and how much security work we’ve left for later.
This goes much deeper than leaked env variables or API keys.
You can protect your secrets properly and still inherit a vulnerability from the software underneath your app.
We’ve seen it with PHP, Firefox, OpenBSD and FFmpeg.
It makes me wonder how much we’ve missed, and how much security work we’ve left for later.
This goes much deeper than leaked env variables or API keys.
You can protect your secrets properly and still inherit a vulnerability from the software underneath your app.
0
0
0
87
2026 shifted the game for indie devs
The quantity game alone feels less viable now
Distribution is still an advantage, but keeping users means solving a problem deeply enough
agents like Codex and Claude Code can already handle tasks that some standalone products charge for
the moats I’d build around:
- proprietary data that improves the product and is hard to recreate
- workflows deeply connected to a business’s systems and history
- network effects where each new user makes the product more useful
- trust earned in an industry where mistakes are expensive
people will still pay for SaaS that saves them time, runs reliably, handles permissions and compliance, and has someone responsible when things break
The quantity game alone feels less viable now
Distribution is still an advantage, but keeping users means solving a problem deeply enough
agents like Codex and Claude Code can already handle tasks that some standalone products charge for
the moats I’d build around:
- proprietary data that improves the product and is hard to recreate
- workflows deeply connected to a business’s systems and history
- network effects where each new user makes the product more useful
- trust earned in an industry where mistakes are expensive
people will still pay for SaaS that saves them time, runs reliably, handles permissions and compliance, and has someone responsible when things break
1
0
1
104
Mohammed S. retweeted
I made Almirah, a little cabinet that lives on your Mac’s desktop.
Drop files, screenshots and scans on it.
Drag them back out when a form asks.
Passport and ID scans go in a locker that opens with Touch ID.
One payment, yours to keep.
Drop files, screenshots and scans on it.
Drag them back out when a form asks.
Passport and ID scans go in a locker that opens with Touch ID.
One payment, yours to keep.
11
3
34
1,311
Last year, the model felt like the main moat
Now choosing between them feels more like a trade-off between output and price
I use GPT and Claude every day
Opus 5.5 performs better than Sol 6.1 for me
Astra gives me better output than Opus, but Opus gives me more for the money
So I don’t have one clear winner
That’s where I think harnesses start to matter more
The tools and integrations become another reason to choose a product and keep using it, even when another company has a better model
Now choosing between them feels more like a trade-off between output and price
I use GPT and Claude every day
Opus 5.5 performs better than Sol 6.1 for me
Astra gives me better output than Opus, but Opus gives me more for the money
So I don’t have one clear winner
That’s where I think harnesses start to matter more
The tools and integrations become another reason to choose a product and keep using it, even when another company has a better model
0
0
0
83
I think a lot of people are already doing this
They just ask AI now for things they used to open
another app for
Canva → make me a simple graphic
Grammarly → fix my grammar
QuillBot → rewrite this paragraph
Jasper → help me write this blog post
Copy. ai → write a few ads for this
Google Translate → translate this message
remove. bg → remove the background
Otter → transcribe and summarize this recording
Excel → analyze this spreadsheet
Google Search → explain this to me
More and more of their work happens in the same chat
They just stop opening the other tools
They just ask AI now for things they used to open
another app for
Canva → make me a simple graphic
Grammarly → fix my grammar
QuillBot → rewrite this paragraph
Jasper → help me write this blog post
Copy. ai → write a few ads for this
Google Translate → translate this message
remove. bg → remove the background
Otter → transcribe and summarize this recording
Excel → analyze this spreadsheet
Google Search → explain this to me
More and more of their work happens in the same chat
They just stop opening the other tools
Software is dead
It's now the era of personal software and everyone just builds what they need whenever they need it I think
0
0
0
85
As AI writes more of the code, I think the best programming languages to build apps with are Python or TypeScript.
Not because AI is always better at writing them; benchmark results depend on the model and task.
But because these languages already have libraries for most of what you need, instead of writing your own library, testing, validating and fixing it.
The future of building is using what already exists instead of building everything from scratch.
You still need to test how those libraries work in your app.
Not because AI is always better at writing them; benchmark results depend on the model and task.
But because these languages already have libraries for most of what you need, instead of writing your own library, testing, validating and fixing it.
The future of building is using what already exists instead of building everything from scratch.
You still need to test how those libraries work in your app.
0
0
0
55
Congrats on the launch @ClimStefan the video looks so interesting.
Aaaand we're live on Product Hunt!!
Would genuinely love your support, check the tool out and leave a comment.
I've prepaired a special offer for you with this special ocasion:
There's a 20% off for first two months, if you use PRODUCT_HUNT at checkout.
Check the launch here: producthunt.com/products/ranks…
1
0
1
91
Mohammed S. retweeted
Aaaand we're live on Product Hunt!!
Would genuinely love your support, check the tool out and leave a comment.
I've prepaired a special offer for you with this special ocasion:
There's a 20% off for first two months, if you use PRODUCT_HUNT at checkout.
Check the launch here: producthunt.com/products/ranks…
Would genuinely love your support, check the tool out and leave a comment.
I've prepaired a special offer for you with this special ocasion:
There's a 20% off for first two months, if you use PRODUCT_HUNT at checkout.
Check the launch here: producthunt.com/products/ranks…
6
2
14
642
Mohammed S. retweeted
I used to drain my 20x plan’s weekly limit in 4 days on average
According to the math:
20x plus ÷ 4 days = 5 plus weekly allowances per day
If the $500 plan gives 25x the usage of Plus:
25x plus ÷ 5 plus per day = 5 days at the same usage rate
That’s only one extra day before I hit the weekly limit, for $300 more per month
The $500 plan is absolutely not worth it .
According to the math:
20x plus ÷ 4 days = 5 plus weekly allowances per day
If the $500 plan gives 25x the usage of Plus:
25x plus ÷ 5 plus per day = 5 days at the same usage rate
That’s only one extra day before I hit the weekly limit, for $300 more per month
The $500 plan is absolutely not worth it .
Imagine paying $500/mo today for what you used to get for $200/mo yesterday lol @OpenAI
0
1
0
86
What I don’t like about GPT models and Codex is how painfully stubborn they can be.
I’m explicitly giving Codex permission to approve certain actions on my behalf, but it keeps refusing, which stops my automated workflow.
How do you deal with this?
I’m explicitly giving Codex permission to approve certain actions on my behalf, but it keeps refusing, which stops my automated workflow.
How do you deal with this?
0
0
0
68
I used to drain my 20x plan’s weekly limit in 4 days on average
According to the math:
20x plus ÷ 4 days = 5 plus weekly allowances per day
If the $500 plan gives 25x the usage of Plus:
25x plus ÷ 5 plus per day = 5 days at the same usage rate
That’s only one extra day before I hit the weekly limit, for $300 more per month
The $500 plan is absolutely not worth it .
According to the math:
20x plus ÷ 4 days = 5 plus weekly allowances per day
If the $500 plan gives 25x the usage of Plus:
25x plus ÷ 5 plus per day = 5 days at the same usage rate
That’s only one extra day before I hit the weekly limit, for $300 more per month
The $500 plan is absolutely not worth it .
Imagine paying $500/mo today for what you used to get for $200/mo yesterday lol @OpenAI
0
1
0
86
Big day for OpenAI with big updates, but overall I was disappointed.
The $200 plan is moving to half its previous usage allowance, measured in equivalent API spend. They’ve also introduced a $500 Pro plan. The new plan supposed to be $300, considering what $200 used to get us.
This seems to pave the way for more price increases. I wouldn’t rule out a $1k plan offering 30x the usage of Plus.
It seems we won’t have the luxury of using AI as freely as we used to.
Don’t lock yourself into one model where every price increase affects you. Try different models and invest in open weight models.
The $200 plan is moving to half its previous usage allowance, measured in equivalent API spend. They’ve also introduced a $500 Pro plan. The new plan supposed to be $300, considering what $200 used to get us.
This seems to pave the way for more price increases. I wouldn’t rule out a $1k plan offering 30x the usage of Plus.
It seems we won’t have the luxury of using AI as freely as we used to.
Don’t lock yourself into one model where every price increase affects you. Try different models and invest in open weight models.
Hi,
Tomorrow we are re-opening the Pro $200 subscriptions to new subscribers, but together with it we are also changing how we calculate the usage for it. In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan.
Now that it's said, let me explain why this is happening and why you will still get more work done than if you were on the Pro $200 subscription one month ago.
(a) We didn't want to compromise in other ways and are committing to not reintroducing the 5h limit, so that you can fully use the weekly usage when you want.
(b) On the subscription, we guarantee that over time you always get more work done and with an increasing level of quality. This means that you will continue to get more value per dollar spent as a result of models getting more efficient and us passing down the improvements in the form of API price reductions.
(c) We don't want to put an incentive on ourselves to artificially inflate the API list prices to make it look like you are getting a lot (and workaround it through discounts, etc). Instead we want to continue to both rapidly reduce prices and increase capabilities of models on the API. This week we introduced GPT-6 Sol and GPT-6 Luna at 50% of their previous price. Over time, we see prices go low enough that it makes sense for most to buy usage as needed without there being a significant gap between what you get in a subscription and what you get in the API for a dollar spent.
(d) Tomorrow, we are adding more things to the subscription that won't draw on the usage, I won't reveal what that is yet.
I wanted to be transparent before all the big announcements tomorrow. Lots of new exciting things are coming to the subscriptions that will make it super compelling, but I wanted to make sure to share this change ahead of time so you can all understand it before we shower you with good news.
Codexingly,
Tibo
0
0
3
286
AI now not only builds demos, but can also build production mobile/web apps in days.
The indie hacker movement that encourages building and shipping multiple apps is not viable anymore. Anyone can build the same app even faster than you.
The game has now shifted to building your product deeply, solving a real problem, and focusing on marketing it.
The indie hacker movement that encourages building and shipping multiple apps is not viable anymore. Anyone can build the same app even faster than you.
The game has now shifted to building your product deeply, solving a real problem, and focusing on marketing it.
9
0
7
434
In first half of 2025, I was satisfied with the $20 claude/gpt/cursor plans.
In second half of 2025 a $20 plan wasn't enough anymore, so I upgraded to $100 claude plan, and it was more than enough to do most of the work.
In 2026, the $100 plan wasn't enough anymore, so I upgraded to the $200 codex or claude plans.
Recently I started to notice that my $200 plans tokens are burned in 3-4 days in average (thanks to resets otherwise I would be cooked). I smell the higher plan is coming soon!
In this way, the plans prices are getting inflated, if opensource models are not catching up, then AI subscriptions won't be afforded by anyone.
In second half of 2025 a $20 plan wasn't enough anymore, so I upgraded to $100 claude plan, and it was more than enough to do most of the work.
In 2026, the $100 plan wasn't enough anymore, so I upgraded to the $200 codex or claude plans.
Recently I started to notice that my $200 plans tokens are burned in 3-4 days in average (thanks to resets otherwise I would be cooked). I smell the higher plan is coming soon!
In this way, the plans prices are getting inflated, if opensource models are not catching up, then AI subscriptions won't be afforded by anyone.
BREAKING 🔥: An upcoming $500 ChatGPT Pro Max plan will likely be powered by @cerebras infrastructure.
> So far, "Fastest Work and Codex" are the only changes in the plan description compared to ChatGPT Pro.
> Usage limits could still be higher as well 👀
* The plan price is $500; T.
0
0
2
177
I use frontier models like Astra, Fable, and Grok daily for work and personal projects.
I noticed recently the performance gap shirinking, and I’m finding it harder to choose a clear winner.
For example, I was fixing a race condition where an authenticated user was reading an old token. I tried several times with Astra, but the fixes still weren’t stable.
Then I opened a fresh conversation with Opus 5.5. I explained the problem, what I wanted to achieve, and what I’d already tried. Its first fix worked as intended when I checked the output and tests.
I’ve had the reverse happen too, which is why I haven’t found one model that works best for everything.
I noticed recently the performance gap shirinking, and I’m finding it harder to choose a clear winner.
For example, I was fixing a race condition where an authenticated user was reading an old token. I tried several times with Astra, but the fixes still weren’t stable.
Then I opened a fresh conversation with Opus 5.5. I explained the problem, what I wanted to achieve, and what I’d already tried. Its first fix worked as intended when I checked the output and tests.
I’ve had the reverse happen too, which is why I haven’t found one model that works best for everything.
0
0
2
136