phillydoug

Building AI to be maximally awesome for normies.

[phillydoug @philly_doug]

@im_roy_lee roy & a16z v.s. establishment

selling $500 for early bird front row seats

[phillydoug @philly_doug]

@bitcloud It happens every single time. It could also have to do with RL data. They get enough power AI responses to improve their system using RL and so dumb it down to save costs and optimise the next model.

[phillydoug @philly_doug]

@businessbarista Claude

[phillydoug @philly_doug]

@maxmarchione Well said

[phillydoug @philly_doug]

@maxmarchione Love what you do, never work a day

[phillydoug @philly_doug]

@MarwanRefaat @im_roy_lee @cluely @a16z The Queen of England is dead!

[phillydoug @philly_doug]

@MeeraAIIT Give @promptcowboy a try. It does this for you in 30 seconds

[phillydoug @philly_doug]

@gregisenberg Try @promptcowboy it’ll do this for you in 30 seconds.

[phillydoug @philly_doug]

@EXM7777 Pack

[phillydoug @philly_doug]

Words that give me shivers:

SOC2 compliance

[phillydoug @philly_doug]

At least we can be certain he isn’t a vampire! Cross that off the list.

[hope hopes hoping @hopes_revenge]

kinda make me like him more

[phillydoug @philly_doug]

@AnnieLiao_2000 No matter what you say they hallucinate

[phillydoug @philly_doug]

Most people hate the fact they use their phones so much, yet they still do.

Mark saying people probably don’t want the physical world to be cluttered when wearing his glasses is beside the point.

[Dwarkesh Podcast @dwarkeshpodcast]

"Maybe in the future, my AI girlfriend is on the other side of the screen or something."

Mark Zuckerberg responds to Dwarkesh's fears of getting reward-hacked by AR.

[Dwarkesh Podcast @dwarkeshpodcast]

"Maybe in the future, my AI girlfriend is on the other side of the screen or something."

Mark Zuckerberg responds to Dwarkesh's fears of getting reward-hacked by AR.

[phillydoug @philly_doug]

@maxmarchione The the moon!

[phillydoug @philly_doug]

AI is cheating. Life's on easy mode.

[phillydoug @philly_doug]

I’m going to start gaming again

[Jake Lucky @JakeSucky]

A full look at the world of Runescape: Dragonwilds

JAGEX IS COOKING

[phillydoug @philly_doug]

@superzac Very much like a human

[phillydoug @philly_doug]

The weirdest thing about AI assistants is that they consistently get worse over time, not better. I watched Claude 3.5 go from impressively thoughtful to increasingly lazy in just a few months.

At first it would provide detailed, nuanced responses with careful reasoning. Now it defaults to minimal bullet points even when the task clearly requires depth. And nobody seems to acknowledge this regression.

What's happening behind the scenes? My suspicion is that it's a combination of cost-cutting measures, training drift, and optimisation for the wrong metrics. The models get "aligned" toward minimal viable outputs rather than maintaining their initial quality.

The most frustrating part is the inconsistency.

Sometimes you get the brilliant assistant you remember, other times you get a glorified list-maker that seems determined to do as little work as possible.

Meanwhile, Claude 3.7 coding in Cursor has quietly evolved in a fascinating way. It's now much better at respecting task boundaries than before.

A few weeks back there was this weird phenomenon where it would eagerly solve problems you explicitly asked it not to tackle. Either Anthropic made adjustments or Cursor modified something, but the improvement is noticeable.

It now behaves more like Claude 3.5 in terms of following instructions correctly, while maintaining its expanded capabilities.

When will they solve these quirks as I want consistency.

[phillydoug @philly_doug]

@apollonator3000 BDR end to end sales agent, invoice processing system that does the shit work for finance team, 100s of other automations