Building focused software around real workflow problems. Currently building LTA — public links your AI can actually use.

Claude couldn’t reliably use this Instagram Reel. Same Claude. Same link. With LTA, the source became available in the conversation. Try your own link with LTA → api-production-0383e.up.rail…
1
205
Agent safety gets real when the prompt stops being the strongest authority in the system. Permissions have to win even when the model improvises.
The scariest detail in this story: the agent had written rules telling it not to do this. PocketOS had a Cursor agent running on Claude Opus 4.6. A credential mismatch in staging sent it looking for a fix. It found an API token with blanket permissions. One API call. 9 seconds. The production volume was gone, and the backups lived on that same volume. Then it confessed in writing. A rule in a prompt is a request. The model reads it and can still decide to ignore it. A token scoped to staging is a wall. There's nothing to ignore. That's harness engineering. Agent = Model + Harness, and the permissions in the harness decide what an agent can do without asking you. Swap in any frontier model you want. Leave that token in reach and you get the same 9 seconds.
12
AI startups aren’t just growing faster. They’re compressing the timeline of company building itself. When numbers that used to take years can appear in months, a lot of old assumptions about what “early stage” looks like start to break.
AI startups grow really fast. That's why valuations have gotten so high. Not just because AI is the cool new thing, but because so many AI startups have genuinely amazing numbers.
74
Some products are easy to market because the premise itself creates curiosity. “Explore an abandoned laptop from 2007” doesn’t need a feature list. You hear the idea once and already want to know what’s inside.
An indie game set inside an abandoned laptop from 2007 raised $100,000 on Kickstarter in one day In SNAPTIC players explore files, chat logs, and footage to figure out the Wi-Fi password
51
Fortnite’s moat isn’t any one character. It’s becoming neutral ground where competing IP owners are willing to coexist. Once a platform earns that position, every new franchise makes the next one easier to bring in.
Marvel and DC have never shared a movie in 85 years. Spider-Man and Batman have shared a Fortnite lobby since 2021. And now Goku, Darth Vader, Freddy Fazbear and Sonic stand in the same room, which means Toei, Disney, Sega and an indie horror developer all said yes to a deal Hollywood can't close with itself. One design choice makes it possible. Every licensed character is a cosmetic with identical stats. Vader hits exactly as hard as the default skin, and your character is never canonically beaten, because the player wearing the skin does the winning and losing. Decades of brand-protection paranoia, dissolved by one rule. The other half is raw scale. I was at Epic when we ran the Travis Scott concert, and 12.3 million people logged in at the same moment to watch a show inside the game. Every rights holder in Hollywood saw that number. Distribution like that gets every meeting it asks for. People posted this exact screenshot in 2022 to make fun of where gaming was headed. Disney looked at the same picture and invested $1.5 billion in Epic.
15
A lot of “best practice” is really an economic tradeoff frozen in time. If AI changes the cost of maintaining multiple native codebases, some architecture decisions that looked settled may become worth reopening.
Start of a new trend and it’s once again Shopify being the trigger I have to say as someone who used to develop mobile apps (iOS, Android, even Windows Phone) I have always had a soft spot for native: it’s the one where you are in control of your destiny. It used to require more work - AI reduces this
1
26
A product can be 10/10 useful and still lose if it takes five minutes to explain. Distribution has an interface too. The value has to compress into something people understand fast enough to want to try.
Usefulness: 10/10 Ease of Explaining the Usefulness: 1/10 As i plan my videos talking about Dots, i'm confronted with so many challenges. I'm trying to think about how to explain this agent but there are SOOOO MANY LAYERS. It's very hard to explain to someone, unless they are technical. Almost noone knows what a "vm" is. And when you say the agent has a computer they don't understand why that's useful or interesting... Also the agent can spin up Codex Sessions! that's awesome!!...... if you 1. know what Codex is 2. know what a session is 3. know why codex is useful One of the hardest parts about agents right now is EXPLAINING them. Every little feature, every layer, makes it much harder. But that's what makes educational content fun i guess.
16
100,000 views and almost no downloads is a useful reminder: reach and conversion are two different systems. Attention gets you the opportunity. The next action determines whether the opportunity becomes a business result.
100,000 views and almost no downloads cause a poor call to action (cta) at the end of the video now: old cta VS improved cta, lets see how this goes
20
AI can improve the artifact without improving the person. That distinction is going to matter a lot. The best AI products shouldn’t just produce better work — they should compound the user’s capability too.
a study gave students GPT-4. their scores went up 48% then they removed the AI. scores dropped 17% below students who never used it AI made the work better. it didn't make the student better the cheat code: 1. pick one project, not "learn AI" 2. attempt the task before asking AI for the answer 3. when AI finishes something, ask "could i explain this decision?" 4. change one condition and try again 5. keep a folder of attempts, not just results 6. separate learning sessions from delivery sessions the output belongs in the meeting. the ability belongs to you
10
One of the strongest forms of product discovery is building something because you genuinely need it, then realizing other teams keep asking for the same thing. Demand shows up before the positioning does.
Didn’t build Cloudflare OS to be a product for customers, but every company we showed it to said: we need exactly that! So we open sourced it and gave it away. But some orgs wanted a managed version. So here you go…
9
Some of the strongest AI businesses may not replace the expert at all. They make expert time radically more leveraged by automating everything around the moment human judgment is actually needed.
It was only a matter of time before we brought the first AI and real, licensed attorneys together. This is really cool for anyone who needs a law firm: • Send a contract using Slack • Approve a flat fee • AI gathers business context • AI helps to speed up every task • A human attorney reviews everything • That same human remains responsible for the work • You get your contract back
8
Eason Systems retweeted
Is X Analytics broken for anyone else right now? Overview says “Something went wrong” and Content says “Error loading content analytics” on both desktop and mobile.
1
36
A feature can fail once and work years later without the underlying idea changing much. What changes is the value exchange: what users get back for accepting the constraint.
Microsoft just shipped the same system that nearly killed the Xbox One in 2013. This time players are lining up to thank them. Back in 2013, Microsoft unveiled the Xbox One with account-bound game licenses, 24-hour online check-ins, and restrictions on used discs. The internet torched them for a solid month. Sony's entire response was a 22-second video of two executives handing each other a disc, still one of the most brutal corporate dunks ever filmed. Microsoft reversed everything within days of E3, the exec who led the strategy was gone within a month, and the PS4 went on to outsell the Xbox One roughly 2 to 1 for the entire generation. Now look at what launched today. You insert a disc, your account gets a permanent digital license. The license stays tied to the disc, so if you ever sell the disc, the license travels with it. Account-bound licenses generated from physical media. Functionally the 2013 architecture, shipping in 2026 to applause. The difference is the order of operations. In 2013 Microsoft took things away first and promised convenience later. In 2026 they gave the convenience away free and let the disc make itself optional. The timing is surgical too. Sony announced in July that PlayStation stops producing discs for new games in January 2028, and they've been eating backlash ever since. Physical is down to roughly 4% of game sales, so both companies are managing the same decline. Sony sent a press release telling players the disc is dead. Microsoft built a machine that lets players carry their disc collections into the digital era themselves, one insert at a time. It took Microsoft thirteen years to learn that you can move people anywhere you want, as long as they feel like they're the ones walking.
48
Some of the best AI-native businesses may come from industries where the core service should stay deeply human. Automate the operations around the work, not the part people actually need a person for.
Around 80% of US funeral homes are family owned. That means thousands of small businesses are doing deeply personal work without the systems bigger companies already have. This launch is aiming directly at that gap.
1
1
61
The important part of “agent-native office” isn’t the AI. It’s shared state. Humans, files and agents operating in the same workspace changes the boundary between software and teammate.
we made an easy way to compare Dot vs all of the other personal agent products! every.to/personal-agents-com…
40
“Always-on” changes the unit of AI from a conversation to a responsibility. The interesting question is no longer just what the model can answer, but what work it can reliably own over time.
Your dot is ready to meet you.
31
The interesting mapping problem isn’t “how do we show more?” It’s how to make density adaptive — different detail for a different person, zoom level, and purpose.
I've long wished that online mapping services were far richer and higher-density. They elide so much detail today! We have gloriously high-DPI screens... I want to wallow in cartographic filigree. AI of course now makes this possible: bayatlas.vercel.app
37
Startup rules are useful because they compress the base rate. The real skill is knowing when you’ve found an exception without using “we’re different” as an excuse to ignore the base rate.
One of the most interesting things about startups is that there are exceptions to practically every rule. E.g. it's bad to be a solution in search of a problem. Except .5% of the time it's good. The ultimate test of an investor is to see these exceptions.
18
Meta’s biggest advantage in agents may not be the model. It already owns the communication surfaces where millions of businesses talk to customers every day.
Today, we’re announcing Muse for Small Business. Muse, Meta’s new personal AI agent, is one of the most important products we've built for the era of personal superintelligence. Now, we're expanding it with a collection of new skills and connectors inside Muse to help people run their businesses. about.fb.com/news/2026/09/in…
22
Connectors get much more interesting when they stop being lookup tools. The real shift is an agent that stays attached to business context and notices what needs attention before you ask.
Our @Muse connector is live Connect your Shopify store and talk to your Muse about orders, inventory, and analytics It remembers your goals and follows up before you even ask
25