DEV Community: Serghei Pogor

GTA 6 made an estimated $1B before launch, and 89% chose the expensive edition. A founder's pricing lesson

Serghei Pogor — Thu, 12 Nov 2026 19:01:40 +0000

A video game is estimated to have made a billion dollars before anyone has played it. I build small products for a living, and that wasn't even the number that got me.

This is a note to myself and to every founder who has ever hidden their expensive plan at the bottom of a pricing page.

What happened

On November 11, Tech-ish and TweakTown reported a new Sensor Tower estimate. GTA VI digital preorders on PS5 and Xbox have passed about 10.31 million copies, worth roughly $1 billion. The game launches on November 19.

For comparison, Take-Two says GTA V took three days after launch in 2013 to reach $1 billion.

Some honest caveats before anyone quotes me:

These are estimates from an analytics firm, not official numbers. Take-Two hasn't disclosed preorders. CEO Strauss Zelnick only said they are "continuing to reach unprecedented levels."
The figure is gross, before Sony and Microsoft take their cut, and it covers console digital sales only. There's no PC version yet.
Sensor Tower hasn't published its method.

So treat the exact number as approximate. The trend is still hard to argue with.

The number I actually care about: 89%

GTA VI comes in two editions: Standard at $79.99 and Ultimate at $99.99. Back in August, Sensor Tower's Karl Kontus said about 89% of preorders were for the $100 edition. He told reporters, as TechRadar quoted, "I have never seen a 90% premium share before." A typical premium share is around 10-20%.

The new totals still point the same way. $1B divided by 10.31M copies is about $97 per copy, which is very close to the expensive price.

Think about it like a restaurant. Normally most people order the regular menu and a few splurge on the tasting menu. Here, nine out of ten tables ordered the tasting menu before the restaurant had even opened.

Why I think this happened

There's no countdown timer and no "only 3 left" banner. The extra $20 buys some vehicles, missions and cosmetics. That's nice, but it doesn't change anyone's life.

What people are really buying is certainty. They already know they'll play this game for years, so the extra $20 doesn't feel like a risk. It feels like going all in on something they've already decided to love.

That trust wasn't built by a pricing page. It came from more than a decade of shipping things people loved, and even a couple of long delays didn't break it.

Where I've been getting this wrong

I've been a PHP developer and team lead for over 13 years, and I've started more products than I'd like to count. Several failed, a few worked. If I'm honest about pricing, I've almost always been the scared one. I keep the top tier cheap, add a discount just in case, and quietly hope nobody notices the price.

The GTA numbers don't prove I should charge $100. They do show that when people trust you, many of them want the better version, and if you don't offer it, you're leaving that choice on the table for them.

The other side (and it's a fair one)

This is not a playbook for a startup that launched last month.

Rockstar spent years earning this. A new product has no reputation to cash in.
Preordering anything unplayed is a real risk. Some big launches in recent years shipped broken, and early buyers paid full price for it.
An extreme premium share can also mean the cheaper option was designed to look worse on purpose. Customers notice that eventually.

So I'm not saying "raise your prices tomorrow." I'm saying I stopped believing my expensive tier is something to apologize for.

A small check you can do this week

If you sell anything with tiers:

Look at what share of paying customers picks your top option.
If it's close to zero, ask whether it's too expensive, or whether it just doesn't offer anything people want.
If you don't have a premium option at all, try adding one: more support, early access, or a bigger limit. Then watch who picks it.

You might find, like me, that the fear was louder than the data.

Your turn

When you buy something with tiers (games, software, a phone plan), do you usually pick the cheapest, the middle, or the top option? And why?

Anthropic's IPO Lets Regular People Request Shares. Here's What That Actually Means

Serghei Pogor — Thu, 12 Nov 2026 16:00:38 +0000

I've built many startups. Several failed, a few worked, and none of them ever had an IPO page. Anthropic, the company behind Claude, now has one: ipo.anthropic.com. It has a feature that caught my eye, because it's aimed at people like my mom, not hedge funds.

Regular people can ask for shares before the stock starts trading. That sounds simple. It isn't, and I think a lot of people will learn how it works the hard way next week.

The numbers, briefly

On November 10, Anthropic filed an updated prospectus with a price range of $70 to $78 per share. TechCrunch reports that this values the company at up to about $1.4 trillion, with 825 million Class A shares in the offering.

The official page says the price is expected to be set on November 18, with trading on Nasdaq under the ticker ANT expected on November 19.

The part that matters for normal people: Anthropic expects up to 20% of the offered shares to go to retail investors. They'll get them through Charles Schwab, Fidelity, Robinhood, SoFi and E*Trade. In big U.S. IPOs retail usually gets around 5-10%, so this is a real change.

"Requesting shares" is not buying shares

This is the part I'd explain to a friend over coffee.

When you request IPO shares in your broker app, you submit an indication of interest. Think of a waiting list at a busy restaurant where you also promise to pay for the whole table, before you know whether you'll get a seat or what the bill will be.

Anthropic's own page is clear about this:

There is no guarantee you'll receive any of the shares you request.
You should be ready to purchase all the shares you apply for.
Your broker confirms your allocation on the first trading day.

Each broker also has its own rules. CNBC reports that Fidelity requires at least $25,000 in retail assets with them, and Robinhood says it only gets "a limited number of shares for each IPO". With demand this high, many requests will be partly filled or not filled at all.

Brokers also tend to dislike "flipping", meaning you get IPO shares and sell them right away. Some can limit your access to future IPOs if you do. Read your broker's rules before you click.

The part nobody puts in the headline

I like that the door is open. For years the biggest gains in tech went to people who could invest before the bell rang. Ordinary people only got the stock after the first-day pop. Giving retail a bigger slice is fairer, and I don't want to sound cynical about that.

The other side matters too, though.

An IPO is also how insiders and early investors start turning paper value into cash. TechCrunch reports that seven co-founders are each selling roughly 3.4 million shares, about $250 million each at the midpoint. That's normal and not a scandal. It does mean someone on the other side of your trade knows the company far better than you do.

Governance is another catch. TechCrunch notes that founders hold a share class with 50.1% of voting power, and a Long-Term Benefit Trust elects the board majority. If you buy Class A shares, you own a piece of the company but have very little say in it.

Then there are the finances. The official page shows Q3 2026 revenue of $17.0 billion, more than 12x year over year, and a net loss of $23.3 billion. That loss includes a $21.9 billion non-cash charge tied to the company's rising valuation. As someone whose startups lost money at a much, much smaller scale, I find that oddly comforting. Nobody gave me a trillion dollars for it, though.

What I'd do before clicking "request"

This isn't financial advice. I write PHP, not stock picks. But this is the checklist I'd run:

Only request what you'd happily pay in full. Assume you get 100% of it.
Assume you might get nothing, and don't plan your week around it.
Read your broker's IPO rules: minimums, flipping policy, how they allocate (lottery, pro rata, fixed amounts).
Decide your plan for day one in advance. The price can open far above or below the IPO price.
Never put money you need next year into one company, however famous it is.

The page also says the registration statement isn't effective yet, so nothing can be sold until it is. Things can still change before November 18.

My take

This is probably the most "normal person" IPO a tech company this size has ever run, and I think that's good. But a door being open doesn't mean you have to run through it on day one. Some of the smartest people I know will just watch the first few weeks and decide later.

Would you request shares in this IPO, or wait and watch how it trades first? Why?

Ten Lines of AI Code Just Shook the Indie Games World. Where's Your Line?

Serghei Pogor — Thu, 12 Nov 2026 13:00:35 +0000

House House, the small studio behind Untitled Goose Game, published an apology this week. The thing they apologised for was roughly ten lines of code.

I've been writing PHP for over 13 years, leading teams and shipping startups (some alive, most not). Reading this story, my first reaction was a nervous laugh. My second was "wait, what's our policy?" Both reactions deserve a closer look.

What actually happened

Here's the timeline, mostly from Inven Global's write-up and Ars Technica's piece:

Nov 9: 19 jurors for the 2027 Independent Games Festival (IGF) Awards put out a joint statement: they will not vote for any game that "has made use of generative AI in any capacity during its development."
One of the games in question turned out to be Big Walk, House House's co-op exploration game.
Nov 10: House House published an FAQ and an apology. According to the studio, late in development they hit a performance bug in Unity. Shadow calculations at sunrise and sunset grew out of control and halved the frame rate for about 20 seconds. They say Unity told them a proper fix wasn't possible in the current rendering pipeline.
So they described the problem to an LLM chatbot, took its suggestion and rewrote it by hand into about 10 lines, under 0.1% of the codebase. No AI was used for art, music, writing, narrative or voice.
They had disclosed this on their IGF submission, because the form asks about any AI use. Steam's rules didn't require it, so the store page says nothing. The studio also wrote: "Developing our game without an internal AI usage policy in place was a failure on our part" (VGC coverage).

As far as I can find, the IGF hasn't announced a decision on Big Walk's eligibility.

Why I don't buy "any capacity"

For non-developers, here's an analogy. A chef's oven keeps switching off mid-service. They ask someone why, get a hint, and fix the wiring themselves. Did that person cook the meal? I don't think so.

Debugging has always been a group sport. I've pasted error messages into search engines, read forum threads from 2011, and asked a colleague at 2am. Most of my best fixes started as someone else's idea that I bent to fit. If help like that counts, no software has ever been made by one person alone.

There's also a practical problem. Modern code editors autocomplete with AI by default. A rule of "any capacity" is one almost nobody can honestly follow. A rule nobody can follow doesn't produce clean games. It produces quiet developers. Notice who got punished here: the studio that disclosed it.

Why the jurors still have a point

I don't want to wave the other side away. Their main argument isn't really about these ten lines. These models were trained on huge amounts of work from people who never agreed to it, and many artists and writers are watching their income shrink because of it. From that seat, "only a little" is exactly how lines move. First it's a bug fix, then a texture, then a whole level.

Programming is also creative work. I've written code I was proud of in the way a musician is proud of a riff. So "it's just code" isn't a full answer either.

Where I land is a moderate spot. Using AI as a debugging assistant is a different category from using it to generate the creative output players actually experience. That's roughly how Valve's Steam disclosure rules seem to be drawn. Awards are free to set stricter rules. Those rules should still be ones people can follow, and they should be clear before submission, not after.

The practical lesson for any team

The most useful line in the whole story is House House's own: they had no internal AI policy. That's most teams I know, including some I've led.

If you run a team, even of three people, here's what I'd write down this week:

Where AI is fine. For example: debugging help, explaining errors, writing tests, boilerplate.
Where it's not. For example: final art, voice, copy that ships to customers, anything under a client contract that forbids it.
How you track it. A simple note in the pull request or ticket is enough. Then you're never reconstructing history from memory when someone asks.
Who you disclose to. Clients, platforms, awards, publishers. Check their forms before you need them.

It takes an hour. It's far cheaper than writing an apology FAQ.

So, where's your line?

I'm genuinely curious where other developers, artists and founders land on this. AI as a debugging helper is fine, AI nowhere near creative work, or anything goes as long as you disclose it? Which one, and why?

Logitech's 'GTA 6 Day Off' Isn't What the Headlines Say, and Why I'd Copy It Anyway

Serghei Pogor — Thu, 12 Nov 2026 10:01:03 +0000

Logitech is giving its roughly 7,000 employees a day off to play video games. You've probably seen the headline as "day off for GTA 6". When I read the actual details, it turned out to be something a bit different, and to me, a bit smarter.

I've been a PHP/Symfony developer and team lead for 13+ years, and I've run a few startups (some died, a few lived). Over that time I've watched plenty of perks come and go. So here's how this one looks from inside a team.

What Logitech actually announced

According to Fortune's report from November 10, the program is called Logi Game Day. Every employee, from full-time staff to interns, can take one day off to play a game of their choice, any time from now until this time next year.

So it isn't really a GTA day. CEO Hanneke Faber plans to take hers on November 19, which is GTA VI launch day and also the day of a company streaming event. That's where the headline came from. Her line in the announcement: "It's time to retire the stereotype that gaming is a waste of time."

One thing Fortune doesn't say is whether this is extra paid leave or a day taken from existing vacation. Some outlets call it paid. If you work at Logitech, I'd love to know.

The numbers behind it

Fortune cites BambooHR data showing that PTO requests for the GTA VI release date are about 13% higher than on comparable days in past years, and that 18% of salaried workers plan to take launch day off.

So almost one in five people were going to disappear on November 19 anyway. The real choice for a company was never "work or play". It was "do we know about it in advance, or do we find out at 9:15 when the standup is half empty".

Why I think it's a good move

Every team lead knows the big-launch-day pattern. Sudden dentist appointments. A cousin's wedding nobody mentioned before. A "slight cold" that clears up exactly when the game finishes downloading.

When you pretend those days are normal workdays, you teach people that honesty is more expensive than an excuse. That's a bad lesson, and it doesn't stay limited to video games.

A named, official day fixes that cheaply:

People plan it in advance, so you can plan around it too.
Nobody has to invent a fake illness.
It costs one day a year per person. That's less than a single bad hire or one burned-out senior quitting.

It's also flexible. Because it's "any game, any day", the person who doesn't care about GTA isn't left out. They can save it for whatever they actually enjoy.

Why I'm not fully sold

I'd be lying if I called this pure kindness.

First, Logitech sells gaming mice, keyboards and headsets. A CEO telling the world "gaming isn't a waste of time" a week before the biggest game launch in years is also very good marketing. Both things can be true.

Second, one play day doesn't fix a heavy workload. If a team is drowning in deadlines, a single day of Vice City is a painkiller, not a cure. The other 364 days matter more.

Third, perks like this can quietly become pressure. If the team culture is "everyone games", the person who'd rather hike or sleep might feel odd. Keeping it framed as "your day, your choice" matters.

What I'd actually do on a small team

You don't need 7,000 employees to copy the useful part:

Name the day. "Take one day this year for whatever recharges you" beats a vague "be flexible".
Ask for a heads-up, not a reason. A week's notice is enough. Nobody owes you an explanation.
Protect the release calendar. My only rule would be: no deploys on a day when half the team is off. I've shipped on quiet days before, and it's never as quiet as you hope.
Watch the other 364 days. If people treat the play day as an escape hatch, that's a signal about workload, not about games.

The question

I'm curious how this looks from your side of the desk. Would you rather get one official "play day" a year, or one no-meetings Friday every month? Which would actually make your work life better, and why?

United Airlines Just Banned AI Agents From Buying Your Ticket. Here's What That Means

Serghei Pogor — Thu, 12 Nov 2026 07:01:16 +0000

For two years, every AI demo has had the same finale. The assistant finds the cheapest flight, picks a window seat and books it while you drink your coffee.

On November 11, 2026, United Airlines took that finale out of the demo.

What United actually changed

United updated its contract of carriage, the long legal document nobody reads before flying. Per One Mile at a Time's reading of the new text, a new rule bans using AI agents "to make, modify or service a Ticket" or to buy any product or service. A second rule lets United cancel reservations made that way.

United told CBS News it will restrict access "directly within our digital platforms and indirectly via travel agencies and aggregators" for unauthorized third-party personal AI agents. So sending your agent through a travel site isn't supposed to be a back door.

What's still allowed: AI tools can show you United fares and schedules. They just can't finish the purchase.

Bloomberg points out the timing: right before the holiday rush.

United's reason: your password in someone else's hands

United's official reason is security. Sharing account credentials with personal AI assistants "can expose sensitive information, and actions taken by these tools may result in unintended purchases, account changes, or use of MileagePlus miles."

CBS also reports United saying agent use could clash with federal rules, such as the requirement that travelers themselves confirm their luggage has no banned items.

I've spent 13+ years building login flows, sessions and permission systems. This part isn't paranoia. Most consumer agents today work by logging in as you, with your full permissions. There's usually no "can search, can't spend miles" switch. It's like giving a stranger your house keys because they're good at writing grocery lists. Most days it's fine. On the bad day, someone books Denver instead of Dublin.

And when that happens, the angry call goes to the airline, not to the AI company.

The other reason nobody puts in a press release

Henry Harteveldt of Atmosphere Research told CBS that United "is saying they don't want to add yet another, and potentially even more powerful, intermediary." He also noted that flight searches cost airlines money.

This is the part I recognize as a founder. Whoever owns the final click owns the customer. That means the upsells, the loyalty data and the chance to sell you a seat upgrade. Airlines spent decades pulling customers away from travel agents and onto their own apps. An AI agent that books for you puts a new middleman back in the middle, and this one doesn't see the upsell screen.

So I don't think it's security or business. It's both. The security worry is real, and it also happens to protect the margin. I've written policies like that myself. You lead with the reason that sounds best.

Not everyone agrees

United is alone for now. Fast Company reports Delta and American haven't followed and say some AI tools already have booking access. CBS notes Southwest made its inventory viewable to ChatGPT users.

This is the interesting part. We're about to see two models side by side: companies that build official, permissioned doors for agents, and companies that lock the door and point at the terms of service. My bet is the first group wins long-term. Still, I wouldn't want to be the first to find out how a rogue agent handles a 3 a.m. rebooking.

Unenforced rules are another story. Critics ask how United will even know an agent made a booking that came through a travel site. United says its systems can identify booking sources. We'll see.

What I'd do if I flew United this season

Let AI search, but do the last step yourself. Compare fares with whatever assistant you like, then book on united.com or the app with your own hands. That's allowed, and nothing can be cancelled under the new rule.
Don't hand your MileagePlus password to any assistant. Not just because of United. It's good hygiene for every account that holds money or points.
Change your password if an agent already has it. If you connected an assistant to your airline account earlier, change the password and check recent activity.
If an agent already booked a trip, check the reservation. It isn't yet clear how United will treat existing agent-made bookings. A quick look at "My Trips" costs nothing.

The bigger picture for builders

If you're building an agent product, this is your warning. "Log in as the user and click around" is a fragile business model. One terms-of-service update can wipe out a flow, and big companies have every reason to write that update. Official APIs, scoped permissions and explicit user consent are boring, but they survive contract changes.

I'll admit some irony. My AI can refactor a legacy codebase in an afternoon, but it can't be trusted with a window seat.

So here's my question: would you let an AI agent spend your money on a flight, or do you always want to press "Buy" yourself? Why?

48 AI Agents Found a Maya Pyramid. Archaeologists Found It First.

Serghei Pogor — Wed, 11 Nov 2026 21:58:06 +0000

A story went around today that made me laugh, then wince, because I've seen the same mistake in a lot of pitch decks.

What happened

Pietro Miotti, who says plainly that he's not an archaeologist, took public NASA lidar data covering about 1,200 km² of the Maya region in Mexico. Lidar is a laser scan from an aircraft, and you can think of it as an X-ray of the ground under the trees. According to his write-up, 48 AI agents went through the strips "one by one like very patient archaeologists".

They flagged a hilltop with a levelled plaza, an oval pyramid of roughly 8–12 metres, a raised courtyard, a possible reservoir and three platforms that were probably houses. He sent a report to INAH, Mexico's national institute of anthropology and history. It replied that "no registered archaeological site was identified at the location given in your report." He says the whole thing took less than 72 hours from idea to official reply.

The story hit Hacker News today and passed 800 points within hours.

The twist

Then came the update dated 11 November. The site had already been mapped in a 2025 paper by Estrada-Belli and colleagues. The agents missed it while going through the literature. It also wasn't in INAH's register yet, which explains the reply, because a missing database entry doesn't prove nobody has found something.

Credit to the author for posting the correction openly. Plenty of people would have quietly deleted the post.

Why this is a founder story

I've been building products for 13+ years, and the most dangerous slide in any deck is the one that says "no direct competitors". It's almost never true. Usually it means the search stopped too early.

The Maya story is a clean example of that, with AI in the middle. The agents did two different jobs:

Finding something. They were excellent at this. Spotting a pyramid shape in noisy lidar is genuinely hard.
Checking whether it's new. This is where it fell apart.

These jobs need different instincts. Finding rewards enthusiasm, and checking rewards suspicion. When you ask one AI session to do both, the enthusiasm usually wins. One HN commenter put it well: agents tend to declare novelty after reading most, but not all, of the existing work.

I've seen the same thing in ordinary product work. Ask an AI whether an idea exists and you often get a confident "there's nothing quite like this". Then a two-minute search in a startup directory shows four of them, two of which shut down for reasons you'd really like to know.

The other side

I don't want to dunk on this. One person, public data and a few days produced a site interpretation that used to need a funded team and a field season. The method worked. The rediscovery even validates the method, because it found something real that experts had also found.

The failure is narrow and fixable. The AI didn't hallucinate a pyramid. It just didn't check the footnotes.

The check I'd use

If an AI tells you something is new, whether it's a startup idea, a bug nobody reported or a "novel" approach in your codebase, do this before you believe it:

Separate the jobs. Open a fresh chat with no context about how excited you are. Ask only: "Find existing examples, papers, products or competitors for X. Links only."
Make it argue against you. "Assume someone has done this before. Where would I find them?" works better than "Is this new?"
Open the links yourself. Half the value is noticing which sources the AI didn't look at.
Ask a human who lives in that world. One archaeologist, one customer or one person on the forum where the topic gets discussed. Their "oh, that's the thing from last year" is worth a hundred tokens.
Treat database silence as weak evidence. "Not in the register" isn't the same as "doesn't exist". The same goes for "not on Product Hunt" and "no results on page one".

What I take from it

AI makes finding cheap. That means more people will "discover" things, and most of those discoveries will be rediscoveries. That's fine, as long as we stop treating "the AI didn't find it" as proof that it isn't there.

The pyramid has waited more than a thousand years. It can wait five more minutes for a literature check.

Have you ever been sure you were first with an idea, a fix or a product, and then found out you weren't? What gave it away?

Amazon Cut an AI Bill by 98% With a Model That Can't Write a Sentence

Serghei Pogor — Wed, 11 Nov 2026 21:19:32 +0000

Yesterday two Amazon engineers published a post with a headline built for Hacker News: Amazon saves millions by replacing Sonnet 4.6 with Strands Decider. By this morning it had hundreds of points and a very opinionated comment thread.

I've led teams for years and shipped a fair number of startups, some of which are no longer with us. This story touches a habit I recognize in myself, so I want to go through what they actually did, why it worked, and why I'm not fully convinced yet.

What happened

Amazon has a huge internal ticketing system covering customer escalations, capacity requests, alarms and security findings. Every ticket needs one of 20 labels. That job was done by Claude Sonnet 4.6, a large and capable language model.

They replaced it with Strands Decider, a "decision model" Amazon launched about a month ago. It can't generate text. You give it some state (the ticket), a question and a list of options. It returns a probability for each option plus a confidence score. That's all it does.

The numbers from their post:

Cost per ticket: $0.0222 with Sonnet vs about $0.0005 with Decider. That's roughly $22,200 vs $500 per million tickets, which they call "about 44 times cheaper" and a 98% reduction.
Speed: average call time went from 4,215 ms to 897 ms.
Tokens: input dropped from 6,554 to 2,009 per ticket, and output from 170 tokens to 1.
Rollback: a three-way config switch (old model, both in parallel, new model). They say going back takes 45 seconds.

The headline claim is that it "performed just as well" and will "save millions of dollars of token costs in a year."

The lawyer sorting your mail

Here's the plain version for anyone who doesn't write code. Using a big language model to pick one of 20 labels is like hiring a lawyer to sort incoming mail into 20 trays. The lawyer will do it well. The lawyer can also write you a memo about each letter, which nobody asked for, and bills by the hour.

The small model is a trained mail clerk. It can't write memos. It just puts letters in trays quickly and cheaply, and it tells you when it isn't sure.

Why I recognize this

As a team lead, I've watched the same pattern again and again, and I've done it myself. You need something working by Friday. A big model gets you there in an afternoon with a prompt and an API key, so you ship it.

Then it works. That's the trap. Nobody puts "check whether this is now a simple job" on the roadmap, so the lawyer keeps sorting mail for two years.

To me the good part of Amazon's post isn't the model. It's that someone went back and looked.

Why I'm not fully sold

Some fair skepticism, much of it from the HN thread:

They sell this model. It's Amazon's own product on Amazon's own blog. That doesn't make it false, but it's a launch story as much as an engineering story.
No accuracy numbers. The post says the new model "beat production Sonnet on a separate set of our tickets" but publishes no precision, recall or sample size. "Just as well" is doing a lot of work.
"Millions" is a projection. No ticket volume is disclosed, so we can't check the math. One HN commenter estimated it would take tens of millions of tickets a year, which is plausible at Amazon's scale.
Cheaper baselines exist. Commenters asked why not a smaller LLM, or a plain classifier like BERT, which people have used for this exact job for years.
They had to bend the problem. The model has a 4,096-token window. They rewrote every label as a one- or two-sentence criterion and drop the oldest comments on long tickets.

The practical check

The authors list their own limits, and I find them the most useful part. It worked because "the task is a closed choice from a known catalog, the volume is high, and the latency of the old path mattered." If you need explanation text, your catalog changes weekly, or the useful context won't fit in a few thousand tokens, they say a decision model probably can't do the job.

So here's what I'd do this week on any product paying for AI:

List every place you call a big model.
For each one, ask: is the answer really one item from a fixed list? Labels, routing, yes/no, priority?
If yes, check the volume. At a few hundred calls a day, don't bother. Your time costs more than the tokens.
If the volume is high, run the cheap option in parallel first, like they did, and compare answers before you switch.

Big models are great for starting fast. They're just expensive to forget about.

Which would you choose for a high-volume job: a flexible model you never have to retrain, or a narrow one that's 40x cheaper but needs care? Why?

Someone Built a Script That Reports Your Idle RAM to Sam Altman. Here's What It Gets Right (and Wrong)

Serghei Pogor — Tue, 10 Nov 2026 12:01:16 +0000

I've been writing PHP for over 13 years, and my laptop has never had a quiet afternoon. Docker, PhpStorm, a browser with too many tabs, a local database I forgot to stop. So when I read that someone built a script to report "lazy" RAM to the CEO of OpenAI, my first thought was: my machine is safe. My second was that this little joke points at something real.

What actually happened

On October 29, Sam Altman went on the Cheeky Pint podcast with Stripe co-founder John Collison. He said it's "quite terrible" for so much of the world's limited memory supply to sit in personal computers that, by his estimate, are used about 5% of the time (Yahoo Finance). That 5% is his own figure, not measured data.

People with home PCs did not love hearing their hardware called a waste. A few days later, Altman said his words were taken out of context and that he wants everyone to have "tons of memory". He also reportedly changed his display name on X to "Ram Altman" for a while.

Then developer Iain Kay shipped ram-altman, and Cybernews, Neowin and VideoCardz all covered it on November 9.

What the script does

It's a small Bash script for Linux (Windows users can run it through WSL). According to the repo and Cybernews:

It checks your available memory every five minutes.
If more than 10% of your RAM is free, it counts that as "missed" memory.
Once a day, it drafts an email to Altman reporting how much RAM you failed to hand over to a data center.

The author left Altman's email address out on purpose, so nobody can turn it into a spam cannon. VideoCardz calls it what it is: a parody. The repo itself even admits a technical gap. Altman was talking about how much of the time memory is used, and the script measures how much is free at one moment. I appreciate a joke with honest documentation.

The part non-developers should know: free RAM isn't wasted

The one thing I'd push back on, in both the original remark and the joke, is the idea that unused memory is wasted memory.

Think of RAM as the free chairs in a meeting room. If every chair is taken, the next person waits at the door. Your operating system keeps some memory free and uses much of the rest as a quiet cache, so the next app or file opens fast. Modern systems already fill "empty" memory with useful stuff and give it back the moment a program asks for it.

So a home PC that shows 40% free RAM isn't lazy. It's ready.

At data-center scale the math really is different. A rack full of idle servers is real money, and that's the world Altman lives in. Memory got expensive, and AI companies are competing with phone and PC makers for the same chips. I understand why someone running huge GPU clusters looks at a billion half-used laptops and sighs.

But the person paying for the laptop is the one who decides whether it's underused.

A 30-second check you can do today

If you're curious, or thinking about buying more RAM:

Windows: press Ctrl+Shift+Esc, open Performance, then Memory. Look at "In use" and "Available".
Mac: open Activity Monitor and pick the Memory tab. Ignore the big "used" number and look at the Memory Pressure graph. Green means fine, yellow means tight, red means you need more or fewer tabs.
Linux: run free -h and look at the "available" column, not "free".

If your pressure stays green during your normal workday, you don't need an upgrade, whatever any CEO or any script says. If it goes yellow or red every afternoon, more RAM is one of the cheapest speed-ups you can buy (even at today's prices).

My take

I like this kind of developer humor. It's harmless, it's well documented, and it turns an awkward quote into something people actually talk about. As a team lead, I've watched plenty of servers sit idle at 3% CPU while we argued about optimizing them, so I see the efficiency point too.

Still, a personal computer isn't a data center that hasn't been given a job yet. Spare capacity is the reason it feels fast when you need it.

Now I'm curious: when you check right now, is your RAM lazy or overworked, and what's eating it? Why?

Gemini Now Shows You Everything It Remembers About You. Here's the 10-Second Check

Serghei Pogor — Tue, 10 Nov 2026 07:22:25 +0000

Every AI assistant you've chatted with for a few months has been building a picture of you. Where you work, who your partner is, which city you keep flying to, the fact that you swore off carbs in March. Until now that picture mostly sat somewhere you couldn't see.

Today Google started showing it. 9to5Google reports that Gemini is rolling out a Memory summary, and Google's help page describes it as a way to "view, edit, and delete what Gemini Apps understands about you."

I've spent 13+ years building products that store user data, so I have mixed feelings about this one. Mostly good ones.

What actually changed

Gemini already had a Memory setting that lets it use your past chats to personalise answers. What's new is one page where the result of all that remembering is written down in plain words.

In 9to5Google's test, the summary was organised into sections: a snapshot of your core background, location and key relationships, then career and projects, travel and daily gear, media and entertainment, and fitness and lifestyle. Each section had specific bullet points.

Think of it as finally seeing the notes your doctor writes while you talk. You always knew they were writing something. Now you can read it.

Who gets it (for now)

According to Google's help page, you need:

a personal Google account, and to be 18 or older
an active Google AI plan, in the US, in English
Keep Activity turned on, and Memory turned on

It's live on the web first. Google's help page also lists the mobile app, but 9to5Google says it hasn't seen it on Android or iOS yet. If you're outside the US or on the free tier, you probably won't see it today. That's normal for a Google rollout.

The 10-second check

This is the part I'd actually do, whatever assistant you use:

Ask it directly. Type: "What do you know about me?" Google suggests exactly this prompt. On the web you can also go to Settings → Personal Intelligence → Memory summary.
Fix what's wrong. You can edit by talking. Google's own examples are "Add to my memory that I speak Japanese" and "Remove my old phone number from your memory." There's also a pencil icon for bigger edits.
Check the history. A "Memory history" view shows when each detail was added or changed. It's handy if something surprises you and you want to know which conversation it came from.
Or switch it off. In the same Personal Intelligence settings, turning Memory off stops Gemini from learning about you, as Google puts it.

One caution: neither Google's page nor the coverage says whether deleting a line from the summary also deletes the old chats it came from. Treat them as separate. If a past conversation held something sensitive, delete it from your Gemini activity too.

Why I think this is mostly good

The uncomfortable thing about AI memory was never this page. It was the invisible pile of guesses behind it. A system that quietly decides "this person lives in X, works at Y, is going through Z" and never shows you is worse than one that shows you and lets you say "no, that's wrong."

I've shipped features where the data model knew much more about the user than the interface ever admitted. Nobody on the team meant harm. It just never felt urgent to build the "show me what you've got" screen. So when a big company does build it, I want to give them some credit.

The other side

There's a fair counterargument. Seeing the summary doesn't change how much was collected in the first place, and an editable page can make people more comfortable sharing even more. Plenty of people will read theirs and realise they never wanted a chatbot to know their kids' names at all.

And memory really is useful. Not having to explain your job, your stack or your diet every single time is most of what makes these assistants feel smart. It's a barista who remembers your order. Lovely, until they also remember your ex's order.

My own position is in the middle. Keep memory on if you get real value from it, but read the summary once a month like a bank statement, and delete anything you wouldn't say to a stranger.

Your turn

If you've got access, try the check and see what comes back. If you don't, ask your current assistant the same question anyway.

Do you keep AI memory on, or do you turn it off? Why?

Android Auto Now Requires Android 11: How to Check Your Phone in 30 Seconds

Serghei Pogor — Mon, 09 Nov 2026 20:35:12 +0000

Your car didn't change overnight. What changed is which phones it will accept.

Google now lists Android 11 or newer as the minimum for Android Auto, the feature that puts your phone's maps, music and messages on your car's dashboard screen. That means phones still running Android 9 or Android 10 are no longer supported. 9to5Google reported the change this morning, and Google's own Android Auto requirements page now shows Android 11 for both ways of connecting.

Here's what happened, how to check whether it affects you, and what your options are.

What actually changed

Android Auto connects in two ways: with a cable (USB) or wirelessly. Until now, the cable connection worked on older phones running Android 9 or 10.

That's no longer the case. Google's official "Get started" help page now lists Android 11.0 as the minimum for both cable and wireless connections.

This didn't come completely out of nowhere. Last month, 9to5Google spotted early signs that this was coming, including a warning some drivers saw telling them their phone would no longer be supported. If you dismissed a pop-up like that while backing out of your driveway, this is what it meant.

One thing to be clear about: "no longer supported" means Google has stopped promising that it works. Based on the reports so far, it's not clear that every older phone will stop connecting the same day. But once a phone falls below the minimum, it stops getting Android Auto fixes and new features, and you can't count on it to keep working.

The 30-second check

You don't need anything technical for this. On your phone:

→ Open Settings

→ Tap About phone

→ Look for Android version (on Samsung phones, it's under Software information)

Then:

→ 11 or higher: you're fine, nothing to do.

→ 9 or 10: check for a system update in Settings (usually under System or Software update). If an update to Android 11 or newer shows up, install it.

→ 9 or 10 and no update available: your phone has reached the end of the road for Android Auto. It may keep working for a while, but it's on borrowed time in the car.

If you need help setting things up again after an update, Google has a step-by-step setup guide.

Why Google is doing this (and why it's still annoying)

My take: this is a reasonable move.

Android 10 came out in 2019 and stopped getting security fixes years ago. Your car's screen shows your messages, your contacts and where you're going. That's not the place for software that hasn't had a security patch in a long time. Supporting every old version also slows down everything else Google ships for Android Auto.

But the other side has a real point too. Many of these phones still work perfectly fine for calls, music and maps. Telling someone "buy a new phone so your car's map keeps working" is a tough sell, especially when nothing about the car itself changed. And quietly updating a help page isn't the same as warning people clearly. A plain, in-app message for every affected driver would have been the decent thing to do.

Your car is fine. It's the 2018 phone that's getting towed.

Your three options

If your phone is stuck on Android 9 or 10, you have roughly three choices:

A) Keep the old phone until it actually stops working.
The cheapest option. The risk: it might fail at a bad moment, like halfway through a trip to somewhere you've never been. Keeping offline maps downloaded on the phone is a good backup.

B) Buy a cheap second phone just for the car.
Some people already do this: a budget phone that lives in the glovebox and only runs maps and music. It needs to be on Android 11 or newer, and you'll need to sort out how it gets internet for maps and streaming.

C) Upgrade your main phone.
The most expensive option, but it solves the problem completely, and you get security updates for everything else on the phone too. If your phone is old enough to be cut off from Android Auto, it's probably missing other updates as well.

There's no single right answer. It depends on how much you drive, how much you rely on Android Auto, and how close you were to replacing your phone anyway.

The bottom line

→ Android Auto now officially needs Android 11 or newer.

→ Check your version: Settings > About phone > Android version.

→ On 9 or 10 with no update available? Plan your next move before the car decides for you.

What would you do: A) keep the old phone until it actually stops, B) buy a cheap second phone just for the car, or C) upgrade? Why?

Claude Fable 5.2 Launch Offers: Who Gets What and How to Claim Your Free Reset

Serghei Pogor — Mon, 09 Nov 2026 20:06:23 +0000

Anthropic released Claude Fable 5.2 today (November 9, 2026), and with it a set of launch offers for paying subscribers. The offers are useful, but they're split by plan and have deadlines. The main free extra, a "limit reset", also isn't applied for you. Here's what you need to know in plain language.

What is Fable 5.2?

Fable is Anthropic's top-tier model family, meant for the hardest, most open-ended work. According to 9to5Mac's launch coverage, Anthropic says Fable 5.2 improves coding, reasoning and vision, and answers up to 30% faster than Fable 5.1. Anthropic still recommends Opus 5.5 as the "daily driver" for everyday work. Those are the company's own claims, not independent tests.

For developers, the ClaudeDevs account lists the API price at $8 per million input tokens and $40 per million output tokens, down from $10/$50.

Who gets what

The official terms are in the Claude Help Center article on Fable 5.2 launch offers:

Pro plans and standard Team seats

Fable models are included in your plan until November 16, 2026, 11:59 pm PT
Fable can use up to 50% of your weekly usage limit during that week
You get one free reset, which also expires November 16

Max, Ultra, premium Team seats (and premium seats on legacy Enterprise plans)

1 to 3 resets, depending on how long you've been paying (ClaudeDevs):
- under 6 months: 1 reset
- 6 to 12 months: 2 resets
- over 12 months: 3 resets
Usable until March 8, 2027

Not included: the Free plan, standard seats on legacy Enterprise plans, self-serve and usage-based Enterprise. HIPAA-ready Enterprise plans get no Fable resets.

There's one more condition: your paid billing must have been in place before November 9, 2026 at 10:00 am PT. If you just subscribed after reading the launch news, this round doesn't include you.

What a "reset" actually does

Think of your weekly limit as a fuel tank shared by all models. A reset gives back only the fuel you burned on Fable. It doesn't make the tank bigger.

The Help Center gives an example: if you've used 80% of your week (50% on Fable, 30% on other models), a reset brings you back to 30%. If your plan also has a five-hour limit, that one refills completely.

A reset works for all Fable models (5.2, 5.1 and 5). Once you use it, you can't undo it. If you cancel or downgrade before using it, you lose it.

How to claim it (the part most people will miss)

Resets are not automatic:

Open Claude and go to Settings → Usage
Find the "Limit resets" section
Press "Reset for free"

Your usage is restored immediately. You don't have to wait until you hit a limit, but there's little point using it early, because it only gives back usage you've already spent.

To actually use Fable, pick it in the model picker. Your default model doesn't change on its own.

Pro users, that reset expires November 16. It's a coupon that expires faster than milk.

My take: generous, but shorter than it sounds

This is a real perk. Letting Pro users try a top-tier model for a week at no extra cost is a fair way to show it off. Long-time Max and Ultra subscribers also get rewarded for staying with extra resets.

But the "free week" may not last a week. Pro users can only spend half their weekly allowance on it. After that, any extra Fable use is billed to usage credits, if you've turned them on. Check that setting if you don't want surprises.

The fair counterpoint: nobody pays anything unless they've opted in to usage credits. Even a day or two with the strongest model is enough to judge whether it's worth upgrading.

So my advice is to save Fable for the one or two tasks that really need it, like a tricky document, a long piece of code or a hard analysis, rather than using it for everyday questions.

Quick checklist

Were you paying before Nov 9, 10:00 am PT? Then you qualify
Pro: try Fable by Nov 16 and use your reset before it expires
Max/Ultra: your resets last until March 8. Check how many you have in Settings → Usage
Check whether usage credits are on or off

Your turn

Which are you this week: A) trying Fable for a real task, B) saving your reset for when you hit a wall, or C) sticking with your usual model? Why?

We Analyzed 1,000 B2B Cold Email Campaigns. Here's What Actually Drives Reply Rates.

Serghei Pogor — Wed, 04 Mar 2026 06:37:51 +0000

We spent time analyzing 1,000 B2B cold email campaigns.

The findings changed how we think about outreach.

The Biggest Number: 38%

38% of emails across these campaigns never reached a real person.

Not caught by spam filters. Not ignored. Simply undeliverable because the contact no longer existed at that email address.

Job changes. Company acquisitions. Abandoned inboxes. Changed email formats.

This wasn't a messaging problem. It was a data problem.

What Separates 5% Reply Rates from 0.8%

When we segmented campaigns by reply rate, a clear pattern emerged:

High performers (5%+):

Verified contacts within 30 days of sending
Refreshed lists every 60-90 days
Bounce rate consistently under 2%
Segmented by actual buying signals, not just industry + headcount

Low performers (under 1%):

Lists 6-12+ months old without refresh
Bounce rates 5-10%
No systematic verification process
Segmenting on broad demographic data only

Same industries. Same email copy. Same SDRs in many cases. Different data discipline.

The Compounding Problem of Bad Data

High bounce rate doesn't just mean fewer replies. It creates a cascade:

High bounce rate
  -> ESP flags domain
    -> Emails land in spam
      -> Open rate collapses
        -> Reply rate hits zero
          -> Pipeline dries up

Recovering domain reputation after serious spam flagging takes 3-6 months. Teams don't realize the damage until they're deep in it.

The Playbook That Works

# The data quality checklist before every campaign
steps = [
    "1. Audit list age - flag anything older than 90 days",
    "2. Run verification on all contacts before sequencing",
    "3. Check bounce rate from last campaign - above 3% = stop",
    "4. Segment by buying signals not just firmographics",
    "5. Source new contacts from verified providers",
]

Tools like RangeLead.com handle the verification and freshness layer - 97% deliverability, updated monthly, decision-makers not gatekeepers.

The ROI Calculation

1,000 verified contacts at 5% reply rate = 50 conversations
10,000 stale contacts at 0.5% reply rate = 50 conversations (plus damaged domain)

Same output. 10x less effort. No domain damage.

Data quality is not an optimization. It's the foundation everything else is built on.

What's your current approach to B2B list quality? Drop it in the comments.

b2b #sales #outbound #coldEmail #leadgeneration