Guys, you NEED to start exploring alternatives like https://t.co/IYyOUaF6n2.
I’m NOT delusional, I know Opus 5.5 is LEVELS ahead of open weights and I use it too.
But to gradually transition from 100% work being done with Opus to maybe 50% Opus, 50% open weights. I currently have 80+ bots, and yesss the ones that generate landing pages or video editing use Opus, but TRUST ME, you do NOT need it for MOST of the work like checking my emails and drafting replies. Doing most of the low level coding work, etc.
So MANY of my bots are using qwen 27B on a BIG ASS GPU system that I rent via runpod at $0.30/hr. And it bills per second, turns on auto if I need it, and then shuts off.
HARDLY $10/mo in usage.
Try it.
new openai model drops and within an hour we already have "agi achieved" posts and "worst model ever" posts
i'll say it: new models stopped mattering for most builders. what actually breaks my products is the model changing behavior every couple months and me rewriting prompts that worked perfectly yesterday
give me boring and stable over smart and new every single time
dario: "when you control a model and you're deploying it, you have the ability to monitor its usage. you can revoke a user's access"
he said that like it's the reassuring part
that sentence is the whole reason i want open models in my stack. every builder who's had limits cut mid-sprint knows exactly who that control is for
and saying it two days before a ~$1.4T IPO is a choice
Dario Amodei on open source models:
“If the path continues, I think we could get to a very dangerous place.”
“When you control a model and you're deploying it, you have the ability to monitor its usage.”
“You can revoke a user's access. You can change what the model is willing to do.”
“When a model is released in an uncontrolled manner, there's no ability to do that. It's entirely out of your hands.”
“We should confront the advocates of this with that problem and challenge them to solve it.”
He clearly doesn't believe Intelligence should be in the hand of the masses and Sam Altman also saying "RAM" shouldn't be accessible to everyone indefinitely agreed with him.
ebay switched off ads on its own brand name. 99.5% of that traffic showed up anyway, through the free link underneath
in 2026, 8 small online brands ran the same test in 9 experiments. none of them lost a significant amount of sales. so why are you paying google for people already typing your name?
> the case for cutting it
- bing experiments, 824 brands bidding on their own name: the ad adds 2-3% more total clicks and pulls nearly half your free organic clicks into the paid link
- one brand campaign audited in 2026: $36,129 spent, only 10.9% bought clicks organic couldn't get
- same company paused paid search. 13 weeks later organic + direct had won back 65% of the sales paid used to claim
- willpowders, 3-week UK holdout: total traffic flat (-1%), organic +6.6% where ads were off. they cut brand spend 90%
- b2b test: brand ads paused for january, organic CTR on the brand name went 45% -> 65%
> the case for keeping it
- same bing data: when a competitor takes the top slot and you sit out, they grab 15-20% of searchers. with 4 competitors it reaches 42%
- with your ad on top, a competitor below you steals 1-2%
- haus, brand search tests: 82% of brands with 3+ competitors on their name saw real lift. low competition: 35%
- orthofeet paused brand ads in 30% of the US. amazon bid on "orthofeet" and took the sales
- EU sports startup, 6-week holdout: -6.8% sales where brand ads were off. their top-of-page share on their own name fell 96% -> 54% as affiliates moved in
> it's getting pricier
- tinuiti: CPCs on your own brand name +19% YoY vs +3% for non-brand (Q1 2025)
- optmyzr, 20k+ accounts, Q2 2026: median branded CPA $42.59 vs $41.67 non-brand. brand is now slightly pricier
- dreamdata (b2b): brand CPC +34% in 12 months, CTR -29%. ~7% of b2b ad budgets go to defending their own name
> check before you touch the budget
- google ads > auction insights on your brand campaign. who shows up, and how often above you
- your organic rank for your name. google's own data: rank #1 = about half your ad clicks would've come free. rank 5+ = 96% incremental
- search terms report. if "yourbrand login" is a top query, you're paying to send customers to the login page. split existing customers into their own campaign
- competitors bid on "brand + login / free trial / demo / reviews" too. one agency found 5 rivals on a single brand
> how to test it
- geo holdout: brand ads off in 30-50% of regions for 2+ weeks. haus median test: 14 days
- measure total sales across every channel (amazon included). google ads conversions will always say brand is amazing
- google dropped its incrementality test minimum from $100k to $5k. 7-56 days, 28 recommended
- keep a tiny brand campaign alive so auction insights keeps working. haus says it can work at ~10% impression share
- performance max: add a brand exclusion list. in haus tests, excluding brand cut new-customer CAC in every test that tracked it, 40% on average
- search campaigns: brand exclusions now live inside AI Max settings
- a rival puts your trademark in their ad text? file a trademark complaint. google won't stop anyone bidding on your name as a keyword
> the fight
one side: brand ads are a tax. agencies love them because they make ROAS look incredible
other side: pull out and a competitor, an affiliate or amazon takes the top slot and up to 42% of your searchers
my take: brand ads are insurance. price them by who's actually in your auction. nobody there = monitoring budget only. 3+ rivals = defend, with a ceiling
> run this week
- open auction insights on your brand terms
- google your own name in incognito
- schedule a 2-week geo holdout
bookmark this before your next ads budget review
everyone bragging they stopped reading the code has never had a paying customer's data break at 2am.
the model writes the code. you still own what it does. read the diff on anything that touches money, auth or user data. skip the rest, i don't care
"i don't read code anymore" is a flex until it's a postmortem
open source lets you take the code. taking it and then telling everyone else's ADE they're ngmi is a choice.
if your repo literally says "ported from T3 Code" in the comments, maybe skip the dunk posts
ship quietly, credit loudly
hot take: they should.
the people who actually catch rockets stay silent, so linkedin fills up with people who've never shipped anything explaining "leadership."
builders posting about real work is the fix. the people who should log off are the ones who make the thought piece and the job the same thing. https://t.co/32oDy1wlPD
My mother said something about AI that I can’t quite get my head around. What’s odd is that she’s a housewife with no tech background.
Last week, I was learning how to use an AirFryer with her and the sort of healthy stuff I could cook with it on my travels.
It was surprisingly easy. And convenient.
I can even control it from my phone.
(I’m planning to turn this into an AI-powered robotic AGI Air Fryer with voice control via Contextli so that I just tell it what I’m in the mood for and it directly injects it in my veins, but this post is not about that).
Within 2-3 days, I felt confident that I could cook anything I want.
And I could cook it faster, cheaper and more customized to MY specific taste.
I told my mother that now I won’t need to rely on restaurants on my travels.
She agreed, and commented:
“Why not post a tutorial about it on X?
No one will go to restaurants if they knew they could cook themselves with an air fryer”.
paused subscribers come back 54% of the time. a "please stay" discount at the cancel button keeps 22% (churnkey data across hundreds of cancel flows)
and apple just put retention offers inside the iOS cancel screen (nov 4). the cancel page is the most under-built screen in most products
> why people actually cancel
- churnkey, ~3M cancel sessions: budget 33%, infrequent usage 31%. their read: "budget" often hides product frustration
- prosperstack, jan-oct 2026: pricing 32%, no time 11%, temporary or seasonal 10%, couldn't build a habit 9%
- under 1% of those answers named an AI tool. ask it as its own option or it hides under "pricing"
- recurly: 51% of consumers who canceled said they weren't using it enough. 38% say they'd rather pause
> what each save is worth (churnkey)
- discount: +5.1 months. 11% still paying a year later
- pause: +5.5 months, picked 15 seconds faster than a discount
- plan change: +7-8 months. 30% still active at 12 months
- recurly: when pause is offered, 25% pick it over canceling
> the order that works
- one question first: why are you leaving
- "busy" / "not using it" -> pause. default 1 month, let them extend
- "too expensive" -> downgrade, then a small one-time discount
- pause goes before any discount. cap repeat pauses
- seat removals -> offer a handoff. superhuman cut seat removals 15% this way
> your save rate is probably fake
- retainup audited 8 subscription brands in 2026: median save 10.8%, best 19.3%
- one app dashboard said 62.5%. remove people who just closed the tab: 12.5%
- https://t.co/9ZZcvk02Jo math: 30% take the discount, 40% of them leave when it expires = ~18% real save
- https://t.co/9ZZcvk02Jo test, meditation app: reason-matched flow vs blanket 60% off = ~1.5x acceptance, ~4.5x LTV per save
- count a save once the next invoice pays
> setup
- stripe portal: retention coupon ships OFF. settings > customer portal > cancellations
- stripe retention rules: offer by cancel reason, months subscribed, total spend. one retention coupon per customer. their example: "too expensive" + 6+ months = 30% off for 3 months
- stripe pause: pause_collection + resumes_at. pick void or mark_uncollectible. keep_as_draft leaves invoices you can collect for the paused months later
- apple: retention messages in app store connect (text, image, switch plan, promo offer), shown after "cancel subscription". real-time API by request, answer in 700ms or apple shows your default
> the legal line
- california since jul 2025: save offers are fine online only with a "click to cancel" button shown next to the offer the whole time
- amazon's "iliad" flow: 4 pages, 6 clicks, 15 options vs 1-2 clicks to join. the case ended in a $2.5B ftc settlement
- ftc says uber one took up to 23 screens and 32 actions to cancel. uber disputes it
- federal click-to-cancel rule got vacated in 2025, ftc restarted it march 2026. ~30 states have their own laws
> the fight
one side: save offers pay your most annoyed customers to stay, and deal hunters learn to hit cancel every quarter
other side: no flow means someone who wanted a month off has to fire you
my take: one screen, reason first, one matched offer, cancel button visible the whole time. if it takes 4 screens to keep someone, you kept a hostage
> run this week
- export 90 days of cancel reasons. if budget wins, read the free text
- add pause as the first offer
- recount your save rate with next-invoice-paid only
bookmark this before you touch your cancel button
The FAA literally grounded every flight in the US over a nationwide communications outage, and Uber’s immediate response is to charge stranded passengers $100–$200 just to go a few miles from the airport. Profiteering off a national travel crisis isn't "dynamic pricing." It's predatory.
good. watermarking AI text was always a dumb idea.
if i draft a post with AI, rewrite half of it and hit publish, that's my post. my name is on it. i take the heat if it's wrong.
nobody stamps "spellcheck was used" on an email. people who want a label on every AI-assisted sentence just want an excuse to dismiss the idea without reading it. https://t.co/OrZOZkZN1Q
chartmogul looked at 2,500 saas companies. trial-to-paid peaks around day 7, and the median daily conversion rate sits near zero until the last day of the trial
people pay when the clock runs out. which means your trial settings are doing more selling than your onboarding emails
> the timing data
- B2B: ~5% of trialists pay within 6 months. half of those pay in the first 7 days, 70% by day 14, 90% inside month 1. months 3-6: about 3%
- first-week conversion: ~2.5% for B2B, ~16% for B2C
- kyle poyar's estimate: 25-40% of signups hit the aha moment in week 1. so 20-35% get value and still don't buy
> card upfront vs no card (2026 data)
- chartmogul/productled survey, 200 products, jan 2026: card-required trials convert 30%, more than 5x no-card trials. only 20% of trial products ask for a card
- per 1,000 visitors: card trial = 35 signups, 10.5 customers. no-card trial = 45 signups, 3.6 customers. freemium = 90 signups, 5 customers
- first page sage (86 saas cos) points the other way: no card 8.5% x 18.2% = ~15 customers per 1,000 visitors. card 2.5% x 48.8% = ~12
- anyone who gives you the answer without looking at your traffic is guessing
> trial length
- 62% of products in that survey run 14 days. 7 and 30 days are 14% each
- revenuecat, 17k+ apps: 17+ day trials convert 42.5% (median) vs 25.5% for 4 days or less
- monthly plans: 77.5% of payers renew once after 17-32 day trials vs 54.2% after 4-day ones
- revenuecat's own caveat: someone still around on day 25 was always more likely to pay
> the plays
- dual CTA: one company in the survey put "free plan" next to "14-day trial, card required" on the homepage. premium trials +26%
- reverse trial: when the trial ends, people drop to a free plan and keep using it. only 7% of products run one
- auto-extend trials for accounts that did your aha action and didn't pay. let the rest expire
- stripe checkout no-card trial: payment_method_collection=if_required + trial_settings.end_behavior.missing_payment_method (cancel or pause)
> the reminder email (where people get burned)
- visa: reminder with an online cancel link at least 7 days before the first charge after a trial. mastercard: 3-7 days before, for digital trials. send it at day -7 and you cover both
- stripe has a toggle: "send a reminder email 7 days before a free trial ends". trials under 7 days get it at signup. the trial_will_end webhook defaults to 3 days, too late for visa
- stripe adds "* TRIAL OVER" to your statement descriptor after the 10th character (22 char limit). check how your first charge reads on a bank statement
- california: trials over 31 days need a notice 3-21 days before the end. sign up online = cancel online
- netflix tested a day-28 reminder in 2016. conversion fell from 90% to 85%. "forgot to cancel" support calls went to zero. they kept it
- canva tells you upfront it'll remind you 7 days before its card-required 30-day pro trial ends. that line sells the trial
> the fight
one side: card upfront is a dark pattern that farms people who forget to cancel
other side: no-card trials fill your dashboard with tourists who never open the app twice
my take: ask for the card when setup takes real effort or the product touches real data, and make the reminder loud. skip the card when people see value in 5 minutes. and every trial should end in a decision screen
> run this week
- pull trial-to-paid by day. if it all lands on the last day, the deadline is doing the selling
- find accounts that did your aha action and didn't pay. give them 7 more days
- turn on the 7-day reminder. today
bookmark this before you touch your trial settings
@attacomsian i think he did have that plan
but it got maxxed out and so not caring for money, he just used api based pricing with cursor
surprising thing is that he didn't know about t3, locul ai or similar software where he could just add multiple subs
Google just introduced 3 new domain extensions and they're actually pretty interesting for small businesses.
.eat - Restaurants, food brands and culinary businesses.
.fly - Travel agencies, aviation and travel-related businesses.
.here - Local businesses, services and location-based brands.
Google is currently running an Early Access Program. General availability starts November 17, 2026.
So if you've been looking for a memorable domain for your business, these are worth checking out.
A good domain won't build your brand for you. But the right one can make it a lot easier for people to remember what you do.
When was the last time you restarted your mac?
MacOS has a memory leak issue; they haven’t solved it yet
It only becomes noticeable if you run like 10+ chats in parallel (even if idle) after about 20-40 days
Nothing fixed it except a restart for me
(not recommending a restart, just sharing info in case it keeps happenign on your system too)
for 2 years every "a chinese lab trained a frontier model on pocket change and nerfed chips" post got 10M views and a nvidia selloff.
if these fingerprints hold up, the newest ones were trained on blackwell.
compute is still the moat. the "you don't need gpus, just be clever" story was a pitch, and a lot of founders made model decisions off a pitch. https://t.co/EyAkexv7OU
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at https://t.co/YOtzND0Jqk.
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
Introducing Claude Meetings: add Claude to your team’s video call and it builds docs, whiteboards, and other artifacts with you while you’re still talking.
Starting to roll out in beta on Team and Enterprise plans.
Introducing Claude Meetings: add Claude to your team’s video call and it builds docs, whiteboards, and other artifacts with you while you’re still talking.
Starting to roll out in beta on Team and Enterprise plans.
your review widget is the reason nobody trusts your reviews
grey box, wrong font, "powered by" in the corner. visitors smell fake and scroll past
paste your URL, it steals your colours. 10 layouts. see every page it's live on
ugly social proof is worse than none
you ask for the testimonial 15 hours after the wow moment. by then they've moved on.
https://t.co/1Qg1iJvvqc asks inside your app, right when they're happy.