YouTube Transcript:
Anthropic has weird vibes

Skip watching entire videos - get the full transcript, search for keywords, and copy with one click.

Video Transcript

developers love Anthropic and there's

good reasons why. They build models that

make it way easier for us to write code

and cloud code is also really really

good. That all said, I don't consider

them good stewards of the open source

and software development world. It

almost seems like they fell into it and

they're getting away with things that I

don't think they necessarily should.

from cutting off access to everyone from

Windinsurf to OpenAI directly to sending

DMCA requests to people who analyze

cloud code too hard on GitHub to just

generally weird vibes about how they

present themselves and talk about

things. I have feelings about anthropic.

I wish I had more good things to say. I

have found myself in a position where

I've had to defend them a lot in the

past. I've been saying for even longer

that I don't really like anthropic.

Today is going to be the day where I

finally sit down and explain to you guys

all of the things that I don't really

like about Anthropic. Please know that

I'm speaking from experience here. As

someone who paid them $31,000

last month, I have been through it with

Anthropic. If we're going to keep

offering teeth reach out at eight bucks

a month, we need to make some money. So,

quick word from today's sponsor and then

we'll dive right in. Coding's never been

more fun. I can do the parts I like. I

can have AI do the parts I don't. and I

can think about my systems a whole lot

more because we can write so much more

code. Code review, however, seems to be

the cost. Now that we're shipping more

code, we also have to review more code.

And I am so tired of the tedious process

of making sure every syntactic thing is

handled, that we're not doing things

that are going to introduce weird bugs.

And I just want to know what the PR

actually is. Today's sponsor, Code

Rabbit, is here to bring AI to our code

reviews. And it has changed how we do

code review fundamentally. I was so

skeptical going in. I really didn't

think something like Code Rabbit would

be for me and I'm so thankful I tried it

because it has changed how we do code

review. It summarizes your PRs. It

leaves diagrams for what's actually

going on, what happens, where, when, and

why. A change summary of each file and

what was affected in it. But most

importantly, it leaves actual comments

and feedback on your PR that have caught

dozens of bugs that we would have

shipped to production. I can tell you

with 100% confidence that there are at

least 40 bugs that I personally wrote

that didn't make it into T3 chat because

code rabbit caught them. Would my other

teammates have caught them? Possibly.

Decent chance for some, less of a chance

for others. There have definitely been

ones the team missed the code rabbit

hit. But more importantly, the team no

longer has to. Their role in code review

isn't going through line by line making

sure I got my syntax correct anymore.

Code rabbit does all that stuff so my

team can focus on the details. The

coolest thing is it learns how you and

your team work. If it gets something

wrong or does something different, you

can tell it at Code Rabbit, no, we don't

do it that way. You can add rules and

other things, too. You can enforce

specific behaviors you want in your

codebase by putting a rule in for Code

Rabbit, and now it will call that out in

review. Like here, we don't use

process.env. We use the T3 env package

so that we have type- safe environment

variable access. And here, someone made

a change that used process.v, and it

yells at them for it and tells them,

"No, go use this instead." This is the

type of review comment that I feel bad

leaving because I don't want to block

the PR. Code Rabbit will do it for you.

And since it does these reviews so

quickly, there's a good chance that you

will get these comments before anyone

else on your team even starts doing the

work. You are no longer inconveniencing

your team with your [ __ ] code. You're

inconveniencing Code Rabbit. It's an AI.

Who cares? If you want to stop bothering

your team with [ __ ] code reviews and

want to focus on the details that

matter, check them out today at soyv.link/codrabbit.

soyv.link/codrabbit.

These are the core pieces I want to

cover. the weird ways they cut off

access, their bad open source behaviors,

their holier than thou attitude, and the

fact that their pricing is, to put it

frankly, absurd. They're like the one

company that has never really lowered

prices and seems to only be slowly

raising them. But let's start with the

cutting off access to competitors

because this is a recent drama. I'm

going to start with this fun quote from

an OpenAI employee. This is a meme

quote, but you get the idea. As part of

our effort to democratize the fruits of

AI and ensure the singularity is

helpful, harmless, and honest, we've

decided to deny others working on the

singularity, the ability to chat with

our helpful, harmless, and honest model.

This is going to be fun. Historically,

Anthropic has tried to position

themselves as this like really positive,

good, kind, working for the best of

humanity company. And every time someone

does that, I get a little bit skeptical.

And I find in their case the

contradictions are pretty absurd.

There's a recent leaked memo from the

CEO, who by the way used to work at

OpenAI where he said, "Unfortunately, I

think the no bad person should ever

benefit from our success is a pretty

difficult principle to run a business

on, wrote Dario in a note to staff about

the possibility of pursuing Gulf State

investments." Yeah,

this is the way Anthropic feels like it

works right now. They're on one hand

trying to be the cool, nice, thoughtful

company building these models, but in

reality they're yet another business and

if anything of the current AI major

labs, they're the one contributing the

least back. Like sure, they have MCP,

which is a cool standard, but at the

same time, they have Cloud Code, which

is the only one of the modern CLI coding

tools that isn't open source. All of the

others are. Not only is Cloud Code not

open- source, but as I hinted at before,

they have been DMCAing people for trying

to reverse engineer it and share what

they found. And there's a very, very

cringe video where the guy who created

Cloud Code dropped some quotes that uh I

have feelings about.

So, we released Cloud Code in February. Yeah.

Yeah.

So, it's been about a little over 3

months. What's it been like? What's the

reaction been from the community?

Yeah, just insane. Yeah, like it's so

unexpected, you know, but before we

released it, we weren't sure if we want

to release it.

Um, it was this tool that internally is

just it makes our engineers and

researchers so productive and we were

having this debate. We're like, is this,

you know, secret sauce? Are we sure we

want to give it to people cuz this is

the same tool everyone at Anthropic uses

every day.

Yeah, they considered not releasing

Claude code because it's secret sauce

and they don't want to give away the

thing that makes them so productive.

God, it just feels so gross. Like I'm

trying to not be like just straight up

rude about this, but like the the

arrogance necessary to even start to

think this way is absurd to me.

Anthropic almost seems like they know

how weak their edges. I'm going to make

a bold statement. The only reason

Anthropic is relevant right now is their

early lead on tool calling. They beat

everyone else to getting that right. So

if you wanted to build your own tools or

your own products around AI, if you're

building your own agents, if you're

building your own thing that runs in the

terminal, if you're building an AI

editor like cursor, wind surf, or

whatever else, anthropics models were

way better at taking a format to

describe some work. Like if you want to

have it be able to look up the weather,

you tell it do tool name equals weather

value equals some zip code. You tell it

that it can do this to get weather

information. and it will get back the

weather information and it will do it.

Their models had this first and it gave

them a huge benefit because up until

recently, there weren't many other

options that would call tools reliably

at all. Even Gemini 25 Pro likes to tell

you what it's going to do with the tools

and then not do it. And I've

consistently found it get nerfed

randomly with updates so that it gets

even worse when you're trying to use it

for real agentic tasks. I still find the

anthropic models to be the best by far

for this. That said, they're now facing

competition. GLM 4.5 and Kimmy K2 are

the biggest risk that Anthropics ever

faced. These models are really good at

tool calling and they're openweight

which means not only can we use these

for free however we want, we can use

them to generate data to train other

models to be good at tool calling as

well. These are two openweight models

that anyone can go download and use that

do the thing that made enthropic so

special a year ago that only they could

do. 3.5 sonnet was a gamecher and it

took us almost a year to catch up

outside of the anthropic bubble but the

catch-up has happened. There are lots of

other models that can do this now. The

Horizon models that are currently

available on Open Router that are

probably by a big lab handle tooling

really well. Newer OpenAI models like

GPT 4.1 does tool calls really well.

These openweight models do tool calls

really well. They're losing their edge.

They were briefly better at UI too. Like

for a while, the anthropic models were

the ones that were best at Tailwind. I

had Claude for Opus generate an image

studio for my video on the Horizon

models, and it this is what it looked

like. It did a fine job. But when you

compare that to what Horizon generated,

it's a night and day difference. Like

they're about to lose their edge on UI.

And this is a non-reasoning model, by

the way. Horizon Beta doesn't have

reasoning, and it was still able to make

something this much better than the most

expensive model that Anthropa currently

offers. Their edge is going fast and the

way that they've historically operated

as a business has been very protective

of that edge. They claim that companies

like Windsurf and OpenAI are trying to

use their models to make theirs better.

They're trying to steal that secret

sauce, the thing that we just heard

about in the previous video. No, that's

a silly way of thinking about it. In a

world where people can't stop publishing

their research, sharing their learnings,

releasing open weights, and more, it's

kind of insane that Anthropic, the only

company that's done all of this wrong.

Anthropic, the only company that

released their CLI closed source. Both

Google with Gemini CLI and OpenAI with

Codeex, have fully open sourced their

CLIs and even take contributions adding

other model support to it. The day that

Codex came out by OpenAI, there was a PR

file that added cloud models and it was

merged the same day. They're also the

only company that doesn't release open

weight models. I know, I know OpenAI

hasn't, but I have a feeling there's one

coming very, very soon. Hell, by the

time this video is published, it might

already be out. They're trying really

hard to get it right. Anthropic seems to

have no interest whatsoever in that.

They're also the only Blab that never

drops prices or does anything to make

things cheaper. Even Gemini overhauled

their cash layer, which was an immediate

drop in the costs for us. They don't.

There are things that they do that are

good. I don't want to just sit here and

rail on them. Their system cards

anthropics to be credited for

snitchbench. I wouldn't have made that

bench if it wasn't for the detailed

system card they published with Opus and

Sonnet with V4. Those system cards gave

us really good ideas and they were also

heavily misinterpreted by a bunch of

lunatics on Twitter that I had to prove

wrong by making a benchmark that's now

relevant and possibly has research being

published about it. Crazy. They also

kind of make standards. Like MCP, as

much as I hate it, is okay. And there

really wasn't a way for them to do that

without just releasing it as a standard.

But uh yeah, I'm trying to think of

other good things to say and I'm

honestly kind of struggling. I just have

rough feelings. You need to account for

my bias and the fact that Anthropic is

one of the biggest bills that I

encounter at any given time. The amount

of my money that these guys are taking

is absurd. And the discount they offered

us was 5% if we commit to minimum 20k a

month of spend. So, if I send them 60

grand because the minimum three-month

commitment, if I hand them 60 grand,

they'll be kind enough to give me 5% off

until we hit the 20,000 of spend and

then we're back to the normal rate

again. What? It's insane. It's

absolutely absurd. But I want to dig

more into the access cutoff stuff

because I think this is particularly

[ __ ] The wind surf one was extra

funny because the Windsurf acquisition

never even went through. Enthropic was

just being aggressive early because

they're kind of arrogant pricks. like

there was no reason for them to do this.

Their chief science officer Jared Kaplan

was interviewed and asked about this and

he said on stage, "I think it would be

odd for us to be selling Claude to

OpenAI." Windsurf got cut off hard from

anthropic, but there's something really

funny about how that works. Anthropic

models can be used on GCP and AWS. You

go to open router, you go to claude

sonnet 4, you'll see that not only can

you use Google vertex and bedrock for

their models, they're actually faster

and often more reliable. It's at the

point where it's hard to use cloud

models just through the official APIs

because open router will make it way

more reliable because when anthropic

goes down, which believe me, they will

go down. They will go down a lot, you

can fall back on Amazon Bedrock or the

many different vertex deployments

instead. So when Enthropic cuts somebody

off, they're not even cutting them off

from access to the model. They're

forcing them to restrict access to

something like the anthropic infra

directly. So you can still use quad in

the wind surf editor. They're just being

petty. The only reason they would do

this is because they are angry and they

want to vent about it publicly. And the

reason they would do that is they want

people to see them for how they feel,

which is what we're doing here. And

we're going to make fun of them for it

because it's petty, narcissistic,

[ __ ] [ __ ] behavior. And I can't

believe a real billion dollar company is

doing this. These [ __ ] are worth $60

billion and they are arbitrarily

restricting access and calling [ __ ]

stupid shots out on stage because

they're so insecure. It's insane. I

can't believe this is just continuing to

happen. And the reason we're filming

this today is because Anthropic and

OpenAI just confirmed the fact that they

have been cut off from cloud access as

well. Why would Open AAI need access?

They need to run [ __ ] benchmarks.

That's why this is an attempt to keep

OpenAI from being able to run benchmarks

comparing the anthropic models to their

models. But as I mentioned before, they

can just use them from the other

providers. So this is purely petty. This

is purely spite. This is purely pathetic

and they deserve to be called out for

their [ __ ] This is something that

an insecure product manager does at a

company. This isn't something that a big

multi-billion dollar company does. At

the very least, they don't do it without

being made fun of, which is what we're

all here for today because this is

[ __ ] pathetic. The more I read into

this one, the more [ __ ] stupid it

gets. Here is the statement from

Anthropic spokesperson. Claude Code has

become the go-to choice for coders

everywhere. So, it was no surprise to

learn OpenAI's own technical staff were

also using our coding tools ahead of the

launch of GPT5. Unfortunately, this is a

direct violation of our terms of

service. Customers are banned from using

the service to build a competing product

or service, including to train competing

AI models. They're also not allowed to

reverse engineer or duplicate the

service. This is pathetic. This means like

I just realized this means we can't use

cloud code. Apparently is against the

code because we're building T3 chat,

which is a chat app that competes with

anthropic. Are they going to cut off

access to the model for our website,

too? Am I going to have to stop offering

anthropic models because they're that

insecure? Is it really coming to that

point? Are they really that pathetic?

Are they going to cut off Google, too,

because Google has better models? I

can't help myself but poke the bear a

little bit. If you haven't already tried

T3 Chat, we were pouring our heart and

soul into this app. It's really, really

good. For eight bucks a month, you get

access to every model versus 20 bucks a

month for getting access to three. Both

are limited, but their limits are way

less transparent, and their app also

sucks. I have a long video describing

all the ways claw.ai AI is a miserable

thing to use. We built something that we

slaved over the details so it isn't

miserable to use. We support all the

cloud models and all of the other really

good models as well for eight bucks a

month. If you haven't subscribed yet,

I'll give you the first month for $1.

Use code weird vibes at checkout for

that discount. Anyways, like is Google

no longer going to be allowed to offer

anthropic models on the cloud? I

wouldn't be surprised if that happens. I

legitimately think there's a moment

coming up where Google and Anthropic

have a big explosive breakup because

they're now making models better than

Anthropic. This is going to get messy

real fast. Let let it be known that I

called this one first. Anthropic's going

to get way pettier and way messier and

start doing things that are obviously

stupid because the things they're doing

right now, in my opinion, are obviously

stupid. Much like DMCAing somebody

because they're fond of the work you

did. This one's insane. Dave Schumer dug

into the cloud code source and he had a

fun time with this. Again, Anthropic is

really insecure. So, they went out of

their way to try and restrict access to

cloud code source even though there were

already existing open source things when

they released it, including ADER, which

is a really good open- source CLI

agentic coding tool built in Python. So,

Dave was curious how Cloud Code's

implementation differed from ADER. So,

he went to look and got a gigantic messy

minified file from what was grabbed from

the npm package. You can see some sentry

stuff in here. So, they're using sentry

within the CLI. Interesting. But they

did have enable source maps on, which

means when you scroll down, you can find

the source map. Apparently though, an

update was pushed when he took a quick

break. And in that update, the source

map URL string was gone. The folks at

Anthropic realized they made a huge

oopsie and pushed an update that removes

the source map. No matter, I'll go grab

the earlier version. Oh, they removed

that, too. Do you know how hard it is to

get npm to remove an old version of a

package? you pretty much cannot do it.

Someone had to pull a lot of strings in

order to get a different version of a

package taken down from npm. That almost

never happens. So, there's a really

funny blend here of the incompetence to

ship the source map URL, the insecurity

to care enough to offiscate it in the

first place, and the connections and

just kind of bullying behaviors that

they have to actually get it taken off

of npm. That is pathetic. Thankfully,

Dave still had it on his machine, so he

was able to go through the source map

URL and actually break it down and find

all of the source code, confirm that

it's using ink. The animation is just

some ass characters. There's no secret

sauce to leak in the system prompt.

Remember, the secret sauce they were

concerned about. There's jack [ __ ] for

that. Files will probably go out of date

pretty dang quick. It's the Anthropic

team's actively developing the tool.

They're already up to 2.19. This was

from 2.8. So, the reward for publishing

that code was a DMCA strike. Thankfully,

GitHub, unlike Anthropic, is a

relatively transparent company. So, all

the DMCA requests they get get published

on their DMCA repo. So, we can see that

Anthropic did this. And we can see this

is far from the only time they have done

this. They really like to DMCA people

trying to rehost cloud code. Look at all

those repos they took down for their own

mistake. The harsh reality is that if it

wasn't for the recent innovations in

tool calling across other models, we

would kind of be stuck with this and

have to deal with it. Thankfully,

there's been a lot of innovation across

other models catching up to and

surpassing Anthropic. It seems like this

has been a focus point for OpenAI for a

lot of this year. As I talked about a

bunch before in the 4.1 drop video and

the 04 mini drop video, OpenAI seems to

really want to win devs back because it

turns out it doesn't matter how smart

your model is on a benchmark. If

developers aren't using it and they're

not building things around it, you have

a real chance of losing in the long

term. And Anthropic's edge has been that

developers like them and build things

around them. That edge is waning fast.

If they were a reasonable company, they

would recognize that and course correct.

They would open source cla code. They

would drop all their models as

openweight. They would lower their

prices a bit. They'd go on a PR

repairing spree trying to fix all of

these things. The fact that Anthropic is

by far the hardest company for me to

talk to and get information from, I can

get our rate limits increased on Gemini

via a DM or a text message. I don't hit

them on OpenAI because we're already

cleared through that. In fact, OpenAI

hits me up letting me know about things

early. They're awesome. OpenAI has been

so nice to work with as an influencer

and as a builder building on top of

their stuff. Anthropic took I had to

pull so many teeth to get a slight bump

in the rate limits on Anthropic's

infrastructure. It was miserable in

order to get Cloud 4 working on T3 chat

day one.

Wouldn't wish that on our worst enemies

or our competition even. The best I can

put it is what I put in the title here.

I just get really weird vibes from them.

My experience working with Anthropic,

building on top of Anthropic, covering

Anthropic, looking into their practices,

finding all these things. Every time I'm

looking more into Anthropic, it just

feels weird. And it sucks because some

of my really smart friends are working

and they're doing awesome things. Some

of their research is genuinely

incredible. The level of detail in their

system cards is super useful. Their

models are making awesome stuff happen,

but the way they run their company is

cringe at best, problematic or damaging

at worst, and we need to talk about it

more. I'm tired of pretending that

Anthropics models being smart for

writing Tailwind means that we can

forgive them for everything when in

reality they are the worst company in

the modern models market when it comes

to working with them as a developer.

We'll see how long it takes for them to

cut off my access now that I've

published this video. Fingers crossed

that they won't, but at the very least,

we'll get some good content if they do.

I got nothing else to say. Thank you for

Click on any text or timestamp to jump to that moment in the video

Most transcripts ready in under 5 seconds

One-Click Copy125+ LanguagesSearch ContentJump to Timestamps

Paste YouTube URL

Enter any YouTube video link to get the full transcript

Most transcripts ready in under 5 seconds

Get Our Chrome Extension

Get transcripts instantly without leaving YouTube. Install our Chrome extension for one-click access to any video's transcript directly on the watch page.

Add to Chrome — Free

Works with YouTube, Coursera, Udemy and more educational platforms

Get Instant Transcripts: Just Edit the Domain in Your Address Bar!

YouTube

←

→

↻

https://www.youtube.com/watch?v=UF8uR6Z6KLc

YoutubeToText

←

→

↻

https://youtubetotext.net/watch?v=UF8uR6Z6KLc

YouTube Transcript:Anthropic has weird vibes

Video Transcript

Paste YouTube URL

Transcript Extraction Form

Get Our Chrome Extension

Get Instant Transcripts: Just Edit the Domain in Your Address Bar!

YouTube Transcript:
Anthropic has weird vibes