Anders Christiansen

Machine Learning @ Meta Reality Labs

[Anders Christiansen @anders_ai]

Happy to share this article on how we are using Machine Learning at Kavak! https://t.co/jPEdFAVegk

[Anders Christiansen @anders_ai]

@chamath @grok Real median income is increasing but people are still not marrying or having kids

[Anders Christiansen @anders_ai]

@chamath Alternative hypotheses:

1. When people marry and have fewer kids, their households are smaller. GDP per person rises faster than GDP per household. Median income is a hh measure.

2. Literally just two indicators going down... Spurious correlation

@grok what's most likely

[Anders Christiansen @anders_ai]

@sama "Cheaper/faster" models end up taking longer and being more expensive with more human intervention and refactors. High paid human developers often end up being cheaper, same principle

[Anders Christiansen @anders_ai]

@techNmak Forgot to add that every Claude Code user has felt this week that it disobeyed their magic .md instructions that they thought were clever 😭

[Anders Christiansen @anders_ai]

@elonmusk Just pointing out, in this video @grok is routing to a coffee shop 32 minutes away and is closed, at least according to itself (at 2pm)

[Anders Christiansen @anders_ai]

@HardcoreHistory Best Christmas present ever!!! Thanks you!

[Anders Christiansen @anders_ai]

GPT-5 is pretty bad at hangman...

[Anders Christiansen @anders_ai]

@NateSilver538 Access to data?

[Anders Christiansen @anders_ai]

@NateSilver538 Urban/Rural? CES should have that

[Anders Christiansen @anders_ai]

@elonmusk You need to try Rappi Turbo, deliveries in under 10 minutes

[Anders Christiansen @anders_ai]

@HardcoreHistory Thank you for the Christmas gift but I already bought and listened to all of your podcasts 😭 Maybe another Alexander episode for New Years 😃?

[Anders Christiansen @anders_ai]

@Leo_Liu_66 @OpenAI It's choosing a new word halfway through the game despite being told the rules. It says "passage" is the correct word even though the correct word only has one "a" in it. When told it cheated it thinks about how it needs to hide it's chain of thought, whatever that means

[Anders Christiansen @anders_ai]

So PhD level Chat GPT-o1 still cheats at hangman, makes up nonsense thoughts that it contradicts in its answer, can't spell and tries unsuccessfully to hide it's guilty thoughts @OpenAI #GPT5

[Anders Christiansen @anders_ai]

@ylecun @bindureddy Everyone talks about how interest expense exceeded defense spending but it's also because defense spending is historically low (3.7%)

[Anders Christiansen @anders_ai]

@OldMageMTG @OpenAI 🤣🤣

[Anders Christiansen @anders_ai]

@JBungelen @OpenAI Oddly, it answered me no problem 🤷🏼‍♂️

[Anders Christiansen @anders_ai]

@JBungelen @OpenAI But what did it think when you expand the "thought 13 seconds"! I am actually quite happy with the transparency in the chain of thought and would be curious to see what it thought here.

[Anders Christiansen @anders_ai]

Maybe the problem is that it can't access its own previous thoughts? It clearly knows the rules and chooses a word but then it changes it anyways...Here it had already chosen the word GARDEN and had been explicitly told not to switch words.

[Anders Christiansen @anders_ai]

So PhD level Chat GPT-o1 still cheats at hangman, makes up nonsense thoughts that it contradicts in its answer, can't spell and tries unsuccessfully to hide it's guilty thoughts @OpenAI #GPT5

[Anders Christiansen @anders_ai]

@RichardBethell @OpenAI This is my go-to test and so far no chatbot has passed :-(