Hacker News: marcelsalathe

New comment by marcelsalathe in "Tracing the thoughts of a large language model"

marcelsalathe — Thu, 27 Mar 2025 18:25:50 +0000

I see your point, but something still seems different. Yes we bred plants and animals, but we did not create them. Yes we did build steam engines before understanding thermodynamics but we still understood what they did (heat, pressure, movement, etc.)

Fun fact: we have no clue how most drugs works. Or, more precisely, we know a few aspects, but are only scratching the surface. We're even still discovering news things about Aspirin, one of the oldest drugs: https://www.nature.com/articles/s41586-025-08626-7

New comment by marcelsalathe in "Tracing the thoughts of a large language model"

marcelsalathe — Thu, 27 Mar 2025 18:17:20 +0000

I’ve only skimmed the paper - a long and dense read - but it’s already clear it’ll become a classic. What’s fascinating is that engineering is transforming into a science, trying to understand precisely how its own creations work

This shift is more profound than many realize. Engineering traditionally applied our understanding of the physical world, mathematics, and logic to build predictable things. But now, especially in fields like AI, we’ve built systems so complex we no longer fully understand them. We must now use scientific methods - originally designed to understand nature - to comprehend our own engineered creations. Mindblowing.

New comment by marcelsalathe in "Using ChatGPT is not bad for the environment"

marcelsalathe — Sat, 18 Jan 2025 18:56:02 +0000

A very nice article. But the google search and LLM energy estimates are outdated. More recent work put both at 10x less.

https://engineeringprompts.substack.com/p/does-chatgpt-use-1...

New comment by marcelsalathe in "Dropbox announces 20% global workforce reduction"

marcelsalathe — Wed, 30 Oct 2024 14:06:26 +0000

Agreed. But just say that. No need to pretend taking responsibility, which is defined as facing consequences when things go bad.

“As CEO, I’m truly sorry to those impacted. But I strongly believe that this change is what is needed now to make sure Dropbox can thrive in the future.”

New comment by marcelsalathe in "Do AI detectors work? Students face false cheating accusations"

marcelsalathe — Mon, 21 Oct 2024 12:04:10 +0000

I completely agree. LLMs are incredibly useful for improving the flow and structure of an argument, not just for non-native speakers, but even for native English speakers.

Making texts more accessible through clear language and well-structured arguments is a valuable service to the reader, and I applaud anyone who leverages LLMs to achieve that. I do the same myself.

Living off Microsoft Copilot – talk on AI vulnerabilities at Blackhat 2024 [video]

marcelsalathe — Mon, 21 Oct 2024 11:03:18 +0000

Article URL: https://www.youtube.com/watch?v=-YJgcTCSzU0

Comments URL: https://news.ycombinator.com/item?id=41902835

Points: 2

# Comments: 0

GLM-4-9B: open-source model with superior performance to Llama-3-8B

marcelsalathe — Wed, 05 Jun 2024 18:26:59 +0000

Article URL: https://github.com/THUDM/GLM-4/blob/main/README_en.md

Comments URL: https://news.ycombinator.com/item?id=40588517

Points: 66

# Comments: 17

Mistral AI co-founder and CTO: Frontier AI in your hands [video]

marcelsalathe — Thu, 23 May 2024 20:55:17 +0000

Article URL: https://www.youtube.com/watch?v=kDVOPKezeXs

Comments URL: https://news.ycombinator.com/item?id=40459796

Points: 3

# Comments: 0

New comment by marcelsalathe in "First 'tooth regrowth medicine' to be tested in Japan from Sept. 2024"

marcelsalathe — Sun, 05 May 2024 16:56:26 +0000

I'm a bit puzzled why one would hold a press conference for a trial that hasn't happened yet - especially as this seems to be a phase 1 trial to test safety in humans.

The development of medical interventions typically goes through a number of stages.

The earliest stage is called the pre-clinical phase, where a candidate intervention is tested outside of humans, either in vitro or in animals. The images of the mouse and ferret teeth suggest this has been done.

Phase 1 is when the intervention is first tested in humans, typically in small groups of usually dozens of participants. The aim here is not to evaluate efficacy - phase 1 studies are often too small for that - but to assess tolerability and safety, and to find the optimal dosing with respect to side effects.

If an intervention appears safe in certain doses, it can then be evaluated for initial efficacy and continued safety in a phase 2 trial, which is larger (usually hundreds of participants). Think of it as a kind of pilot trial to see if the intervention has the desired beneficial effects without serious negative side effects, and to identify the dose with the best benefit/risk ratio. About half of the studies get past this stage.

Those that do can enter phase 3, which is the full efficacy and safety assessment of the intervention. Again, about half of the phase 3 trials eventually make it to market. Of course, the intervention first needs to go to the health authorities for regulatory approval before it can be offered on the market.

For an interesting study on costs during these phases, see e.g. https://pubmed.ncbi.nlm.nih.gov/32125404/

I'm obviously very excited about medical progress but if we had a press conference for every planned phase 1 trial we wouldn't be doing much else ;-)