Dries

senior software & ML engineer; freelancer

[Geoff Brumfiel @gbrumfiel]

An absolutely astonishing account of Israel's AI-assisted targeting in the @latimes

They called 62-year-old Ahmad Turmus and asked: “Ahmad, you want to die with those around you or alone?” His response, below, is chilling.

Story by @nabihbulos
https://t.co/sBVBNxOMnA

[Dries @driesder11]

@Shpigford This should be possible with OCR to get the name, team, number... from the image. Not easy to get accurate though. There's also a reverse image search on tcdb: https://t.co/6VWUbp3uiQ

[Dries @driesder11]

@Shpigford firefox

[Dries @driesder11]

@anthonynsimon Some benchmark score to compare them, since 1 vCPU can differ a lot

[Dries @driesder11]

@Shpigford Definitely 7 / 0, although I'd pick 6:50 / 0:10 over either if it was a daily choice for optimal health.

[Dries @driesder11]

@Shpigford 3. Batch check (this is assuming the cost is mostly because of the 14k categories you are inputting) 4. Do a categorization on the first hierarchy level (Electrical, Flooring...), then on the rest.

[Dries @driesder11]

@Shpigford Some ideas: 1. use a cheaper LLM to categorize, then verify with a better one (smaller query so cheaper), and only do a second categorization if incorrect. 2. Create a dataset from GPT-4o, then fine-tune a cheaper model on that data.

[Dries @driesder11]

@eshear Did you happen to ask what he does when two 'trustworthy-looking' people present him opposing ideas?

[Dries @driesder11]

@srbaker from Dall-E 3

[Dries @driesder11]

@karsten_frank It doesn't say 'all artifacts and all creatures' so I'd assume the exception counts on 'artifacts and creatures' so it doesn't destroy itself even if artifact.

[Dries @driesder11]

@Shpigford @DetangleAI You have a large waiting list, right? Ask them what they'd use it for.

[Dries @driesder11]

@chrisrickard I'm only just starting so I'm not getting inbound leads yet. Most of my emails are cold and I'm getting responses/calls/meetings from them.

[Dries @driesder11]

@levelsio I usually decrease the learning rate in higher iterations of the model training.

[Dries @driesder11]

@NithurM @lemonsqueezy 30 dollars is a lot on a 1000 payout. On their website: "Lemon Squeezy does not charge any fees to process a payout. However, our payout providers will charge a small fee to process the payout which will be deducted from the total payout amount."

[EFF @EFF]

The EU is considering a proposal that would allow for widespread scanning of user messages. https://t.co/COfUS71hyv

[Dries @driesder11]

@chalmagean I'm at step 3. If anyone needs some help, dm me.

[Dries @driesder11]

@Shpigford Ruby/Rails/Postgres with Devise is what I use. I usually pick Bootstrap but Tailwind is fine too if you know it already

[Dries @driesder11]

@edisonparklive you mean 40s right?

[Dries @driesder11]

@chrisrickard 👍

[Dries @driesder11]

@fuslie Just scrolling on Twitter 😋