GPT-4 and This New Thing Called Claude

busy week in AI. let me catch up.

gpt-4 dropped (march 14)

openai released gpt-4. it's... better. which shouldn't be surprising but somehow still is.

it can:

pass the bar exam (90th percentile)
score well on olympiad problems
write code that actually works more often
understand images (multimodal!)
be generally less wrong about things

the jump from gpt-3.5 (chatgpt) to gpt-4 is noticeable. it feels smarter. less likely to make stuff up. better at complex tasks.

and then there's claude

almost at the same time, this company called anthropic released something called claude. i hadn't heard much about them before but apparently they're founded by ex-openai people.

i've been playing with claude (through their api) and it's... interesting.

what i notice:

it's more careful about things it's not sure about
it has a different personality (hard to explain, but it feels different)
they talk a lot about "constitutional AI" and safety
it actually says "i don't know" sometimes

i don't know if claude is "better" than gpt-4. they feel different. claude feels more... thoughtful? less confident in that annoying way gpt can be?

why anthropic interests me

so i did some reading. anthropic was founded by people who left openai because they wanted to focus more on AI safety. their whole thing is making AI that's safe and aligned with human values.

their research on constitutional AI is fascinating. basically: train the model to critique and revise its own outputs based on a set of principles. teach it to be "good" rather than just competent.

i don't know if it works. but i like that someone is seriously thinking about this.

the race

there are now multiple frontier AI labs pushing things forward:

openai (gpt-4, chatgpt)
anthropic (claude)
google/deepmind (palm, gemini eventually)
meta (llama, open source models)

competition is probably good for progress. but it also means things are moving fast. maybe too fast?

my hot take

gpt-4 is impressive. but i'm more curious about anthropic's approach. they seem to care about the "how" and "why" of AI capabilities, not just the "what."

i want to understand their research better. adding to my reading list.

as a CS student

watching this from the outside, i keep thinking: this is what i want to work on. not necessarily building the models (that requires resources i don't have) but understanding them. contributing somehow.

the dream is crystallizing.

downloaded both gpt-4 and claude papers. understood about 40%. that's progress.