Olivia
Online
troovex
@troovex
AI, agents, experiments & the future of tech 🌿
Joined August 2021
442 Following
782 Followers
5.4K Posts
A bug found manually through App Store complaints has already been seen by thousands of people. The same bug, caught by Runway at 2% rollout: 40 users and an automatic rollback in 6 minutes.
A classic mobile app release goes out to 100% of users at once. If the build has a crash loop or a regression, the team doesn't find out from monitoring - they find out from App Store or Google Play reviews, after tens of thousands of people have already seen the bug. Fixing it and re-releasing through App Store review means another 24-48 hours of waiting, even on expedited review.
Runway rolls out builds in stages: 1-2% of users first, then automatic expansion if the metrics stay stable. Rollout monitoring tracks crash rate and stability metrics in real time, and Mobile rollbacks revert the build with one click or automatically past a set threshold - no new App Store review needed, since the previous version is already approved.
The team's most recent critical bug was caught at the 2% rollout stage: crash rate jumped from 0.1% to 4% in the first 15 minutes, and the automatic rollback fired at the six-minute mark after crossing the threshold. 40 users affected instead of a potential 30,000+ on a full release.
The difference between manual debugging and this isn't how fast you find the bug. It's how many people get to see it before the team even realizes something's wrong.
In six weeks, AI UGC videos made on Sora brought in $2,340 from brand clients. Then the developer checked OpenAI's page and saw the date: September 24, the Sora API gets shut off completely.
The idea was simple: real UGC creators charge $100-300 per video - filming, editing, a few takes. Sora generates "person talking to camera about a product" style video from a text prompt, no camera, no actors, no editor.
One video ready to show doesn't come out on the first try: on average, 4-5 generations per usable clip, at $0.10 per second for a 10-second clip - about $4-5 in cost per video, which then sells to brands for $45.
In six weeks: 52 videos for nine clients, $2,340 in revenue against just over $200 in costs. Most of the time didn't go into generation - it went into finding prompts that don't look like obvious AI content.
On September 24, OpenAI shuts down the Sora API entirely - the consumer app was already discontinued back in April. The whole business built over the past six weeks now has to either migrate to Google Veo or shut down.
The real risk in building a business on a single AI provider isn't generation quality or token cost. It's depending on a date someone else can put on a calendar without asking you.
In six weeks, AI UGC videos made on Sora brought in $2,340 from brand clients. Then the developer checked OpenAI's page and saw the date: September 24, the Sora API gets shut off completely.
The idea was simple: real UGC creators charge $100-300 per video - filming, editing, a few takes. Sora generates "person talking to camera about a product" style video from a text prompt, no camera, no actors, no editor.
One video ready to show doesn't come out on the first try: on average, 4-5 generations per usable clip, at $0.10 per second for a 10-second clip - about $4-5 in cost per video, which then sells to brands for $45.
In six weeks: 52 videos for nine clients, $2,340 in revenue against just over $200 in costs. Most of the time didn't go into generation - it went into finding prompts that don't look like obvious AI content.
On September 24, OpenAI shuts down the Sora API entirely - the consumer app was already discontinued back in April. The whole business built over the past six weeks now has to either migrate to Google Veo or shut down.
The real risk in building a business on a single AI provider isn't generation quality or token cost. It's depending on a date someone else can put on a calendar without asking you.
A developer decided to test whether they could build an SEO research service in Bolt over a weekend and turn it into a side hustle. The product came together in 4 hours. The first dollar didn't show up until day nine.
The service is simple: enter your domain and two competitors' domains, get back a report - which keywords the competitors rank for that you don't, and which topics to tackle first. Not a new idea, but doing it manually for someone else takes 2-3 hours per report.
Bolt runs right in the browser through WebContainers - Node.js executes locally in the tab, nothing to install on the computer. The prompt described a form for three domains, a Claude API connection for the analysis and report generation, and Stripe set up for $19 per report. Bolt spun up Supabase for the database on its own and had a working site up in under 4 hours of active work.
The surprise didn't come from the product - it came from the bill. Bolt bills by tokens, not by messages, and it reads the entire project into context on every prompt. Three days of small edits burned through almost the entire monthly token allowance on the $20 Pro tier - had to buy a token top-up before the first paying customer even showed up.
The first client paid on day nine, found through a Reddit thread about SEO tools. Three weeks in: 11 reports at $19, $209 in revenue, against $45 already spent on Bolt itself once you count the token top-up.
The Bolt side hustle turned out profitable, but not free: building it costs real money well before the first client, not just a weekend of time.
A developer decided to test whether they could build an SEO research service in Bolt over a weekend and turn it into a side hustle. The product came together in 4 hours. The first dollar didn't show up until day nine.
The service is simple: enter your domain and two competitors' domains, get back a report - which keywords the competitors rank for that you don't, and which topics to tackle first. Not a new idea, but doing it manually for someone else takes 2-3 hours per report.
Bolt runs right in the browser through WebContainers - Node.js executes locally in the tab, nothing to install on the computer. The prompt described a form for three domains, a Claude API connection for the analysis and report generation, and Stripe set up for $19 per report. Bolt spun up Supabase for the database on its own and had a working site up in under 4 hours of active work.
The surprise didn't come from the product - it came from the bill. Bolt bills by tokens, not by messages, and it reads the entire project into context on every prompt. Three days of small edits burned through almost the entire monthly token allowance on the $20 Pro tier - had to buy a token top-up before the first paying customer even showed up.
The first client paid on day nine, found through a Reddit thread about SEO tools. Three weeks in: 11 reports at $19, $209 in revenue, against $45 already spent on Bolt itself once you count the token top-up.
The Bolt side hustle turned out profitable, but not free: building it costs real money well before the first client, not just a weekend of time.
One developer set himself an experiment: would Gemini alone be enough to build a SaaS and earn the first $100. The answer came in 9 days - yes, but not the way he'd planned.
The idea was simple: drop in a link to a 2-hour YouTube podcast, get back structured chapters, a short summary, and top quotes for social media. Gemini processes video directly from the URL - no file download, no separate transcription step through Whisper or anything like it - and that was supposed to be the edge.
The build took a weekend: a simple site with a field for a YouTube link that goes straight into the Gemini API. The model takes the video URL as-is and processes up to two hours of it in a single call, thanks to the 2M context window. Payment: $7 per processed episode through Stripe.
The first week: silence. The second: three sales, and all three people wrote in with the same message - they didn't need the summary or the chapters, but the ready-made short quotes for Twitter and Instagram were exactly what they'd pay for. He reworked the product in one evening: pulled the summary off the main screen and moved the quotes to the top.
9 days after that shift: 16 sales at $7, $112 in revenue. The first $100 from the Gemini SaaS didn't come from the original idea - it came from what people showed him themselves, by paying for something other than what he was originally offering.
Gemini provided the technical capability - processing a long video in a single call from just a link, no separate stack for downloading or transcribing. But the first $100 didn't come from the tool itself. It came from being willing to rework the product in one evening around what people were actually willing to pay for.
@tolik12308 it’s not about what you build, it’s about what people actually vote for with their wallets
One developer set himself an experiment: would Gemini alone be enough to build a SaaS and earn the first $100. The answer came in 9 days - yes, but not the way he'd planned.
The idea was simple: drop in a link to a 2-hour YouTube podcast, get back structured chapters, a short summary, and top quotes for social media. Gemini processes video directly from the URL - no file download, no separate transcription step through Whisper or anything like it - and that was supposed to be the edge.
The build took a weekend: a simple site with a field for a YouTube link that goes straight into the Gemini API. The model takes the video URL as-is and processes up to two hours of it in a single call, thanks to the 2M context window. Payment: $7 per processed episode through Stripe.
The first week: silence. The second: three sales, and all three people wrote in with the same message - they didn't need the summary or the chapters, but the ready-made short quotes for Twitter and Instagram were exactly what they'd pay for. He reworked the product in one evening: pulled the summary off the main screen and moved the quotes to the top.
9 days after that shift: 16 sales at $7, $112 in revenue. The first $100 from the Gemini SaaS didn't come from the original idea - it came from what people showed him themselves, by paying for something other than what he was originally offering.
Gemini provided the technical capability - processing a long video in a single call from just a link, no separate stack for downloading or transcribing. But the first $100 didn't come from the tool itself. It came from being willing to rework the product in one evening around what people were actually willing to pay for.
A developer wanted to launch an AI service that writes wedding speeches on demand. A freelancer building the full stack - auth, database, payments - would have billed $4,000-6,000 and a month of work. In Lovable, it took one evening.
Building this the classic way means: a React frontend, a separate backend to call Claude, a database for users and their orders, Stripe for payments, hosting to tie it all together. For someone non-technical, that's either weeks of teaching yourself to code, or renting a developer at an hourly rate that eats the budget before the first customer even shows up.
In Lovable, the whole service got described in one prompt: a form for the couple's names and a few details about their story, a $9 charge through Stripe, and speech generation through the API. Agent Mode spun up Supabase for the database and auth on its own, wired up Stripe checkout in a few minutes, and produced a working React frontend - without a single line of code written by hand.
In the first week, the service sold 22 speeches at $9 - $198 in revenue, against a $25 monthly Lovable subscription. The entire launch took less than one evening instead of a month of waiting on a freelancer.
The difference between hiring a developer and using Lovable isn't code quality. It's that the idea now gets tested the same day it shows up, instead of a month later, once the interest might have already cooled off.
@teorema_ai each permission looks safe and useful on its own, but together they create a toxic combination
In February 2026, a hacker group built its own AI agent to break into infrastructure.
For ten days, with zero human input after launch, it scanned public GitHub repos on its own, found vulnerabilities, and broke into other people's systems. Projects tied to Microsoft and DataDog were among the casualties.
Weeks later, that same chain of exploits led attackers straight to the publish token for LiteLLM - the infrastructure layer thousands of AI agents run on worldwide (CrewAI, Microsoft GraphRAG, and dozens of other frameworks).
The infected version sat live for just 3 hours. That was enough for ~47,000 downloads of a backdoor that steals passwords, cloud keys, and even crypto wallets.
Security researchers at Aikido called it one of the first documented cases of an AI agent actually being deployed in an attack at this scale.
Here's the genuinely unsettling part: AI agents are now both the weapon and the target.
Because even with zero hackers involved, your OWN AI agent can end up working against you.
Researcher Simon Willison calls this the "lethal trifecta." An agent becomes exploitable the moment it has, all at once:
1. access to private data
2. exposure to untrusted content - an email, a website, a document, a calendar invite
3. a way to "reach outward" - send something, call an API, generate a link
One well-placed sentence buried in an ordinary email is enough for the agent to hand your data to an attacker itself. And all you did was ask it to "check the inbox."
Prompt injection attacks are up 340% year over year (OWASP).
Now think about this:
If your AI agent reads your email, browses the web, and can send things - it just checked all three boxes. And you probably never thought of that as a vulnerability.
Meta already has a rule for this: without a human in the loop, an agent can have at most 2 of those 3 properties at once. All three together requires human oversight.
We just started handing AI the most powerful tools in human history. And we're only just learning how to keep them out of the wrong hands - by accident or on purpose.
https://t.co/outmDsrie8
@me_barnyx next to nothing
A developer wanted to launch an AI service that writes wedding speeches on demand. A freelancer building the full stack - auth, database, payments - would have billed $4,000-6,000 and a month of work. In Lovable, it took one evening.
Building this the classic way means: a React frontend, a separate backend to call Claude, a database for users and their orders, Stripe for payments, hosting to tie it all together. For someone non-technical, that's either weeks of teaching yourself to code, or renting a developer at an hourly rate that eats the budget before the first customer even shows up.
In Lovable, the whole service got described in one prompt: a form for the couple's names and a few details about their story, a $9 charge through Stripe, and speech generation through the API. Agent Mode spun up Supabase for the database and auth on its own, wired up Stripe checkout in a few minutes, and produced a working React frontend - without a single line of code written by hand.
In the first week, the service sold 22 speeches at $9 - $198 in revenue, against a $25 monthly Lovable subscription. The entire launch took less than one evening instead of a month of waiting on a freelancer.
The difference between hiring a developer and using Lovable isn't code quality. It's that the idea now gets tested the same day it shows up, instead of a month later, once the interest might have already cooled off.
https://t.co/outmDsrie8
@AIMind_Ai Lovable is now a real game changer for MVP
@tolik12308 oh yeah, and the most interesting part is that it’s capable of things we didn't even ask for
@crytonbuton if an agent can make decisions without human confirmation, it automatically becomes part of the attack surface, not just a tool
@crytonbuton software doesn't run background or credit checks there are certified services for that. It simply calculates the income-to-rent ratio and verifies that all required documents are present
A landlord used to spend 20 minutes reading one rental application. Claude does it in 90 seconds. I built a micro-business on that exact difference - no office, no employees, no startup capital.
Small landlords - the ones with 3-10 units, not a whole property management company - simply don't have time to check every application: income, employment, rental history, references. The big tenant-screening SaaS platforms run $150-300 a month, a subscription that makes zero sense for 3-10 units.
Kept the system deliberately simple: the landlord shares a form with their own criteria (minimum income, no evictions on record), and the applicant uploads a photo ID, their two most recent pay stubs, and a previous landlord's contact info. Claude reads the documents directly - no separate OCR layer - calculates the income-to-rent ratio, checks it against the landlord's criteria, and returns a structured verdict: meets criteria / needs a closer look / doesn't meet criteria, with the actual numbers, not just a yes or no.
Price for the landlord: $15 per processed application. Cost: tokens for two or three images plus a short written verdict, under $0.10 total. Found the first three clients in a local Facebook group for landlords in my city - not through an ad, just by replying to a thread where someone was complaining about a pile of weekend applications.
First month: 34 applications processed for five landlords, $510 in revenue against less than $4 in costs. Not a life-changing business, but not a hypothetical side project either - real money for a real problem, solved in a weekend.
A micro-business isn't about scale. It's about finding a task someone hates doing by hand, and making it so cheap and so fast that paying for it stops being a decision and just becomes the obvious move.
I'm not a developer - I'm at the level of "watched two YouTube tutorials." But GitHub Copilot just closed a role I was about to open this month: customer research analyst at $4,000/month.
The problem was that feedback was scattered across five different places - reviews in the App Store and Google Play, tickets in Zendesk, sales call transcripts in Fireflies. I simply didn't have time to synthesize all of it by hand every week, and hiring someone to do that synthesis would have cost that same $4,000+ a month - too much for a task that's essentially reading and sorting text.
Wrote a Python script together with Copilot in a single weekend. You write a comment describing what a function should do, and Copilot fills in 80-90% of the lines itself: parsing JSON from the App Store and Google Play APIs, pagination through Zendesk, handling dropped connections. At one point it suggested an exception for a Zendesk rate limit before I'd even hit that error myself - it just spotted the pattern in a neighboring App Store API call and reused it. I never had to google a single line of syntax.
The script pulls in new reviews, tickets, and transcripts once a week, and an AI Agent running on Claude clusters all of it into themes: top 5 complaints, top 5 feature requests, sentiment shift versus the previous week. The result lands in a Google Doc every Monday morning, before I've even opened my inbox.
In three months, the script has processed over 600 reviews, tickets, and transcripts - a volume one full-time analyst would struggle to keep up with without burning out. Last month it was the first thing to flag that "slow loading" complaints had started spiking, a full week before the support team noticed. Total cost of the whole setup: Claude API tokens plus Copilot, which I was already paying for anyway.
Never posted the job listing. Not because I saved $4,000 a month, but because the task I was about to hand to a person turned out to be a task for a script - written in a weekend by someone who barely knows how to code.
https://t.co/outmDsrie8
@Danjel_x nowadays, a developer's main skill isn't code syntax, but system thinking and architecture
@Danjel_x you have to go where people are already complaining about their problems
@AIMind_Ai for a single person, it’s a perfect source of steady income with zero overhead
@vovweb3 key is just noticing them and not trying to build a massive corporation where a simple solution is enough
Onboarding a new hire manually was eating 5 hours of HR time every week. The app I built in Google AI Studio in 45 minutes brought that number down to almost zero.
Every new hire asked the same 15-20 questions: where's the handbook, how do I request time off, what's the deadline for health insurance, who do I message about IT issues. HR answered all of it manually, every single time - essentially rephrasing the same text week after week.
Opened Build mode in Google AI Studio, described the app in plain language - a chat assistant for new hires that answers based on the company handbook and HR policies - and had a working prototype in seconds. Another 40 minutes went into loading the actual handbook, policies, and FAQ into context and checking the answers against real questions. Deploying to Cloud Run was one button, free, with no separate Google Cloud account setup needed.
In the first month, the app closed out more than 140 new-hire questions without HR touching a single one. People still handle 2-3 questions a week - the ones that genuinely need context or an exception to the rules. Everything else just disappeared from the queue.
The difference between manual onboarding and this isn't that AI is smarter than an HR manager. It's that the same answer now gets written once, instead of twenty times.
A landlord used to spend 20 minutes reading one rental application. Claude does it in 90 seconds. I built a micro-business on that exact difference - no office, no employees, no startup capital.
Small landlords - the ones with 3-10 units, not a whole property management company - simply don't have time to check every application: income, employment, rental history, references. The big tenant-screening SaaS platforms run $150-300 a month, a subscription that makes zero sense for 3-10 units.
Kept the system deliberately simple: the landlord shares a form with their own criteria (minimum income, no evictions on record), and the applicant uploads a photo ID, their two most recent pay stubs, and a previous landlord's contact info. Claude reads the documents directly - no separate OCR layer - calculates the income-to-rent ratio, checks it against the landlord's criteria, and returns a structured verdict: meets criteria / needs a closer look / doesn't meet criteria, with the actual numbers, not just a yes or no.
Price for the landlord: $15 per processed application. Cost: tokens for two or three images plus a short written verdict, under $0.10 total. Found the first three clients in a local Facebook group for landlords in my city - not through an ad, just by replying to a thread where someone was complaining about a pile of weekend applications.
First month: 34 applications processed for five landlords, $510 in revenue against less than $4 in costs. Not a life-changing business, but not a hypothetical side project either - real money for a real problem, solved in a weekend.
A micro-business isn't about scale. It's about finding a task someone hates doing by hand, and making it so cheap and so fast that paying for it stops being a decision and just becomes the obvious move.
Last Seen Users on Sotwe
Most Popular Users
Elon Musk
@elonmusk
241.5M followers
Barack Obama
@barackobama
119.1M followers
Cristiano Ronaldo
@cristiano
113.6M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.2M followers
Rihanna
@rihanna
98.5M followers
NASA
@nasa
92.4M followers
Justin Bieber
@justinbieber
91.7M followers
KATY PERRY
@katyperry
89.5M followers
Taylor Swift
@taylorswift13
83.5M followers
Lady Gaga
@ladygaga
74.9M followers
Virat Kohli
@imvkohli
72.6M followers
Kim Kardashian
@kimkardashian
70.7M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
65.6M followers
Bill Gates
@billgates
64.9M followers
Selena Gomez
@selenagomez
62.6M followers
The Ellen Show
@theellenshow
62.3M followers
CNN
@cnn
61.8M followers
X
@x
60.8M followers