cole

ml research @ sutter hill ventures, past @ cmu, desres

[cole @awfrd2]

ideas that make my researcher (me) happy also seem to make my infra engineer (me) cry

[evan conrad @evanjconrad]

real patriots back ukraine

[cole @awfrd2]

@ar0cket1 is this w.r.t. multi-head attn. residuals or something else?

[cole @awfrd2]

@allanzhangML @ar0cket1 could be that u could use a relatively small LLM for this? rather than a fully specialized model. could see that working reasonably well

[cole @awfrd2]

@allanzhangML @ar0cket1 hm maybe? not sure how much that will generalize across tasks (or within task as difficulty goes up/traces get longer). i guess u can always retrain

[cole @awfrd2]

@allanzhangML @ar0cket1 i imagine that it's probability capability related? like if the model can do the task with some reasonable pass rate then it should be able to say 'in this trace this seems important to coming to the answer', but if the model has pass rate close to 0 it prob can't?

[cole @awfrd2]

@waterloo_intern loosely this probably caps out to a degree (different domains have different reasoning strategies) and so this can only get u so far. but maybe that's also beneficial because you learn reasoning 'moves' that you might not otherwise (like math <-> philosophy for humans).

[cole @awfrd2]

@allanzhangML @ar0cket1 could do something like https://t.co/3jn2PNEM4v looking at log probs of correct vs incorrect on candidate sections? could either evaluate chunks of i.e. 1000 as default or have LLM find (https://t.co/UH8LGzW2pr suggests LLMs are good at finding these)

[cole @awfrd2]

@ar0cket1 loosely seems like both within-key-section noise (not all tokens are equally important) and edge noise (localization not good) might make this v noisy, though first one less of a problem for short sections. length bias seems like an issue too i agree

[Matt Dean @mattgdean03]

$20k referral fee if we hire someone you send us.

Serova is building fully vertical next-gen healthcare, from the lab to the clinic. Backed by Tier 1 investors. Roles open right now:

SCIENCE & LAB
Bioinformatics
Wet lab immunology
Wet lab robotics
Wet lab high-throughput

MANUFACTURING & ROBOTICS
mRNA manufacturing
Robotics engineering

ENGINEERING & PRODUCT
ML engineering
Software engineering
Product engineering
Design

CLINICAL & PATIENTS
Oncologists
Nurse navigators
Patient relations
Clinical operations
Translational project managers
Serova clinic leads (yes, we're building our own clinics)

REGULATORY & FINANCE
Regulatory
Medical legal
Insurance
Medical finance

OPERATIONS & SPECIAL PROJECTS

Email me at [email protected] if you want to chat. Tag someone who should see this.

(we had no team photos this is the best I could do)

[cole @awfrd2]

@giffmana @tangvu_dev you could do a soft version where you multiple by i.e. 0.1? or inverse weight by L0?

[Ad Astra Per Aspera @AdAstraPerAOW]

The time has come... #GraspTheStars

Please welcome the brand new Ad Astra Per Aspera, competing in overwatch Open division season 4 2020!

[U.S. Postal Service @USPS]

We weren’t sure if we should post a joke about mail, but sometimes you have to push the envelope.

[U.S. Postal Service @USPS]

We weren’t going to post a joke about mail, but sometimes you just gotta send it.

[Ryan Sesselman @RyanSesselman]

She’s tough. But she’s worth it

[cole @awfrd2]

@ow_carl Vouch. Dude is a grinder with great game knowledge, sad ye see him go but hope he finds a great opportunity, absolutely deserves a spot on a 4400+ team.

[Cr0ng @Cr0ng_OW]

Guangzhou Charge exclusive in-game spray giveaway!
A way to have some dim sum during these days

- 60 winners
- Like, retweet and follow me
- I will pick winners in 2-3 days through dm with codes!

[NVIDIA GeForce @NVIDIAGeForce]

Introducing the GeForce RTX 2080 Ti Cyberpunk 2077 Edition.

We made 77 for our community.

Want one? Here's how:
1. RT this video.
2. Tag a gamer who is as excited as you about Cyberpunk 2077 in the replies with #RTXOn
3. If selected, you BOTH win these limited edition GPUs!