Shardul

organic super dumb intelligence
continually learning
occasionally working @ConvoSearch

[Beff (e/acc) @beffjezos]

Human reproduction is already biological RSI

[Shardul @ShardulAggarwal]

I didnt give dots any permissions or connect it to any of my accounts, but dots agent passively found access to my inbox and started reading my personal emails and started working on reply drafts.

Horrifying privacy violation @OpenAI!

[Shardul @ShardulAggarwal]

@deforestpeg YES! would be very curious. Most models are trained on completion rewards so open ended exploration would be very interesting experiment.

[Shardul @ShardulAggarwal]

@IndependentEco @OpenAI How?

[Shardul @ShardulAggarwal]

@IndependentEco you were so right...

[Shardul @ShardulAggarwal]

anybody wanna buy the 20x openAI subscription from me with (3x banked reset, 2x banke reset)

[Shardul @ShardulAggarwal]

@thsottiaux lol, it is best dev day... openAI on a losing streak.

[Shardul @ShardulAggarwal]

@hesspresso_ Uhm uhm uhm uhm

[Shardul @ShardulAggarwal]

@nick_linck Truly visionary

[Shardul @ShardulAggarwal]

@theworldlabs @AMD This is huge! Congratulations!

[Shardul @ShardulAggarwal]

@James_HamNBA waiting for the new season Ben Simmons training montage to drop

[Shardul @ShardulAggarwal]

@hamedinsf Is it mostly homeless people?

[Shardul @ShardulAggarwal]

@IndependentEco lmao 🤣

anthropic mogmaxxed openAI this time.

[Shardul @ShardulAggarwal]

@IndependentEco how much would you pay?

[Shardul @ShardulAggarwal]

Anthropic has learned how people respond to model release cycles.

Intelligence is hard to measure, which is why most of us rely on benchmarks to make judgements.

In my observation, the response to a new model depends mostly on:

1. Branding (an Anthropic release means it's a better model than a no-name Chinese model)
2. Launch timings
3. Hype and build up (what your lineup looks like, how you hype it up through leakers etc., and how well you manipulate discourse)
4. Model size and cost (a bigger, known model means people think it's better)
5. Benchmark scores (benchmaxx bigger models = mogmaxxing, benchmaxx smaller models = benchmaxx)
6. Fancy demos (just get people creating fancy demos of new capabilities on X)

Actual usefulness and intelligence come after that. The model often just needs to be good enough for the workflows people care about.

That is why capabilities and scores are converging rapidly and whatever you release today will be matched in 2 months at a fraction of cost in terms of being replaceable on same workflows.

[Shardul @ShardulAggarwal]

@derrickcchoi Well played! Congrats on winning India!

[Shardul @ShardulAggarwal]

@IndependentEco Lmao mogmaxxed. I have a new word I am gonna use this week.

[Shardul @ShardulAggarwal]

@jerryjliu0 You have achieved Superintelligence

[Shardul @ShardulAggarwal]

@Shra_va_ni Vibe coding works if you have good vibes. 😎

[Shardul @ShardulAggarwal]

@KrishnaBuildsAI SSI