Gary Illyes

Depending on what side of LinkedIn you're on, you're going to run across JEV, hyped as the new shit of LLMs. It is not. It's the marketing hype of discriminative models, induced by some folks discovering that frontier LLMs really shouldn't be used for classification tasks (for a bunch of reasons)

LLMs really really suck at being random. Much like hallucinations, that isn't a bug. It's how autoregressive models work, it's by design. Testing an LLM with a coin flip tells you nothing about its actual "intelligence" or reasoning capability; it just reveals human training data bias. I like 7.

Google's crawlers don't parse JSON, they just download things. Parsing anything is done by, for example, indexing in case of Search.

IP addresses don't have a "geographic location". They typically default to whatever the registrant's location is, or they're defaulted to some location based on BGP anycast routes. Think about addresses like 1.1.1.1 or 8.8.8.8. Where do you put these on a map? They're technically everywhere...

Gemini might have had enough of my nonsense. That period is ruthless...

brb. need to blow up our snowman www.sechselaeuten.ch

Gemini just told me it's a pleasure working with me, so, yeah... I'm def putting this on my resume

Processing crawl complaints, sometimes I wish I could throw some hosters under the bus. Why would you redirect your client to us for "excessive CPU usage" when the WHOLE MONTH there were about 1000 fetches from us from the WHOLE DOMAIN?!?!

Gary: I can't get this stupid ai to do anything right. Gary's prompts: Found a second cottage cheese, yes we did, precious! A cold lumpy 115g of it... and so salty, it stings our throat! What we do with it my love? We mix it with the nasty leaves? Garden-salads? Or do we eat it raw, just as it is?

A couple weeks ago we accidentally posted on LinkedIn the wrong description for a podcast episode. The comments under that post seem to be a great way to find automated accounts. www.linkedin.com/posts/google...

I'm perpetually confused about people liking our ramblings in these podcast episodes. youtu.be/JpweMBnpS4Q

ARE YOU TELLING ME MY LYNX BROWSER NEEDS TO BE UPDATED?!?!

Who comes up with a name like "Punch, the monkey"?! No wonder everyone is bullying the poor thing

Ten points for... well, not effort for sure

Genius tech bro move: Inventing the paragraph, Calling it a chunk. #showerthoughts

#showerthought What if the moltbook is not a demonstration of sentience, but a reflection of humans' (and bots') online shenanigans.

#showerthought LLMs rarely reply with "I don't know" to a prompt because we, humans, rarely ever admit online that we don't know something.

The coding agent I'm testing just suggested that we send temperature readings in Celsius to an LLM to make conversions to Fahrenheit, instead of making the conversion on the device, because ultimately we want reliability. So, yeah...

Posit 1: browsers don't particularly care about the status code of the parent frame, they'll happily display wherever content the server sent. Posit 2: crawlers very much care about the status code and they'll drop anything with non-content/transition status. #MisuseEverything

"Cursor CEO warns vibe coding builds ‘shaky foundations’ and eventually ‘things start to crumble’" fortune.com/2025/12/25/c...

If we migrated the googlebot.json file to a new location with a 301 redirect, would that break your fetches? Or the fetcher would just follow the redirect developers.google.com/static/searc...

www.google.com/search?q=67 I'm so confused...

Subtwee... subblue: using something as an API doesn't make it an official API. developers.google.com/search/blog/...

Ask me how it's going at the Shenzhen SEO Conference.

Why do crawlers generally have a per-resource bytesize limit, like Googlebot's 15Mb? The internet: news.ycombinator.com/item?id=4467...

24 CPU cores working for 35 hours to train a model, 1% left to completion, and accuracy and loss values look phenomenal. Windows update restarts the PC...

Allegedly intelligent systems that allegedly excel at summarization requiring a manually updated file that essentially says what a site is about feels as intuitive as a sign explaining that the provided toilet paper is not meant to be used for giftwrapping the toilet.

SEO was claimed dead in 1997. web.archive.org/web/19981206...

O.M.G. PEOPLE! UPDATING THE COPYRIGHT YEAR IN THE BOTTOM OF YOUR PAGES IS NOT A SIGNIFICANT UPDATE, NO NEED TO UPDATE YOUR LASTMOD

Just finished listening to the latest episode of SciShow Tangents, which means I've listened to every single one of the Tangents episodes @hankgreen.bsky.social et al. put out on the public web. AMA open.spotify.com/wrapped/shar...

Nothing says better "I want to wipe out my site from search" than these bad boys. Ok HTTP status codes do. And "noindex". Connection issues? Ok there are many things.

🔥 It's kinda fun watching people lose their shit about the Reddit blackout and blaming centralization, when Usenet (and more recently Tor) had the same problem, but with decentralized communities. You woke up and the community was not there anymore. Then Verizon and co fixed that with centralization

you can't spell cat without a...

"site: example.com" =/= "site:example.com"

While the crypto folks were pushing web3 on the blockchain, there was quite extensive work being done on quantum communications. In fact China demonstrated that entangled photons can carry information to and from a satellite. Somehow, investing time and money in this seems more sensible?

The mitigation on Google's side is now in place so at least Google services are back online in the Philippines. Would love to hear your theories about how an incident like the one described in the article I linked to is mitigated by multi-cluster service owners.

Looks like a cable cut in South/South-East Asia is severely affecting internet users in the Philippines: https://unbox.ph/news/major-internet-outage-in-the-philippines-google-facebook-down/

Check your firewalls and access rules: if you want to use the inspection tools in Search Console, you must allow Google-InspectionTool to access your site (and your clients' sites, if you're a hoster or such)

Should we add something new to sitemaps?