David Ferris @d4ferris

technical staff @openai previously founded @phonicAI (YC W20, acq) @google @uber @uwaterloo @playground_ai @southpkcommons. 🇺🇸/🇨🇦 dferris.me Toronto / SF Joined January 2022

Tweets

587
Followers

624
Following

1K
Likes

1K

David Ferris @d4ferris

2 weeks ago

@WillManidis There's a real market for actual shit-talking

0 0 2 294 0

View Details

David Ferris @d4ferris

3 weeks ago

We should put the datacenters in Saskatchewan

0 0 3 82 0

View Details

David Ferris @d4ferris

3 weeks ago

@dev_kunjadia Reply asking if they’d be willing to price your equity to the pre-raise valuation

0 0 1 355 0

View Details

David Ferris @d4ferris

3 weeks ago

@DamiDina So sick!! btw I'm kinda bullish on visual reasoning, there's an interesting company @vlmrun who seem to be the only ones taking this seriously rn run by @spillai who I met at SPC

1 1 2 225 1

View Details

David Ferris @d4ferris

3 weeks ago

@damidina These kinda of vision failures emerge in many agentic loops that uses VLM-as-judge, but especially those in design

1 1 1 86 0

View Details

David Ferris @d4ferris

3 weeks ago

@damidina tldr; VLMs are (still) blind to tons of really glaring design issues

1 1 1 73 0

View Details

David Ferris @d4ferris

3 weeks ago

@DamiDina Maybe white-paper is a better name for it-- just wrote up a bunch of ideas in design/vision/AI and a few experiments I ran.

1 0 1 41 0

View Details

@DamiDina Great insight. literally wrote a paper about how vision is a bottleneck: dferris.me/files/reconstr… some of it is lossy vision encoders, mostly it's is a lack of vision tasks in VLM training data. Totally solvable btw, just lower priority rn than coding/reasoning/et. al

2 1 2 189 1

View Details

David Ferris @d4ferris

4 weeks ago

It feels like a pretty significant anthropological loss that we don't have these early models anymore. I hope that the weights are open-sourced someday so we have complete documentation of the takeoff.

0 0 2 286 0

View Details

David Ferris @d4ferris

4 weeks ago

@tokifyi bro your site is down

1 0 1 74 0

View Details

David Ferris @d4ferris

4 weeks ago

startup gtm tier list, iykyk

2 1 6 370 2

View Details

David Ferris @d4ferris

4 weeks ago

@VivekxK of course that's true

0 0 2 185 0

View Details

David Ferris @d4ferris

4 weeks ago

2019 narrative, you missed the inversion. It's now much easier to join a startup than it is to join FAANG. Still lots of unmotivated people in big tech, but relatively few unskilled ones. The talent floor at startups is *far* lower.

altra @catboosted

4 weeks ago

Crazy how fast working at FAANG went from high status to professional mark of death

30 31 946 382K 67

4 1 30 6K 6

View Details

David Ferris @d4ferris

4 weeks ago

the snake game has me in an absolute chokehold

Y Combinator @ycombinator

4 weeks ago

Pops (@pops_fyi) lets anyone create and play AI games. Describe your idea, and Pops turns it into a fun mini game to remix, share, and play with friends. No code required. All you need is an idea. Short-form software. Congrats on the launch, @alonzuman! ycombinator.com/launches/QPl-p…

45 29 268 119K 163

2 0 2 271 1

View Details

David Ferris @d4ferris

4 weeks ago

This is because "startups" are no longer (a) that risky and (b) a monolith. There is a long tail of low quality, small companies who raised some $. Startups went from the moonshot bet for high-agency engineers to the backup plan for mid ones. There are still some great startups, but It's a minefield.

0 0 10 336 1

View Details

David Ferris @d4ferris

4 weeks ago

I'm pretty skeptical of these kinds of calculations, they seem true only if work is highly monotonous. The value from difficult, white-collar labour is inherently pretty spiky and benefits from random moments of inspiration and connection. Also humans are not computers that simply turn on and off- fun begets culture begets motivation. That said, most meetings *are* too long and poorly run. Big fan of "leave the meeting if it's not useful" cultures.

1 0 1 46 0

View Details

David Ferris @d4ferris

4 weeks ago

btw if you want style LLM outputs it's much more effective to *demonstrate* that style in the system prompt than to state it explicitly. "Remember — no em dashes!" is inane human crafted prompt => humanized output