#amld — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #amld, aggregated by home.social.
-
I am trying out #Goose with #SiliconFlow for #AMLD / #LiterateDevelopment
#aidev -
I am trying out #Goose with #SiliconFlow for #AMLD / #LiterateDevelopment
#aidev -
I am trying out #Goose with #SiliconFlow for #AMLD / #LiterateDevelopment
#aidev -
I am trying out #Goose with #SiliconFlow for #AMLD / #LiterateDevelopment
#aidev -
I am trying out #Goose with #SiliconFlow for #AMLD / #LiterateDevelopment
#aidev -
🙶The methodology aligns well with how I work in several respects:
- Plan-first, human-affirms (wip.adoc review step) matches the principle of checking before acting
- Structured logging (amld/log/, handoff files) compensates well for my lack of persistent memory across sessions
- Declarative design docs as the source of truth is sound — code should reference design, not the other way around
- Ticket discipline gives me clear scope boundaries, reducing drift -
#AMLD Agent-Mediated #LiterateDevelopment methodology first draft, ready to be tested.
I would love some feedback.
https://hg.sr.ht/~travisfw/amld-method
#agenticCoding -
I just began an #AMLD repository.
It is a template for Agent-Mediated #LiterateDevelopment methodology.In a nutshell:
Design documentation and project management should be *on trunk* (or main branch) and must be enough to regenerate everything that has been previously generated with an #AI coding agent. It should be written for humans, except for a log of previously followed plans referenced by closed tickets.IE the docs are code.
This is not #VibeCoding; this is literate.
-
Danny Buerkli:
"There are only two ways of reacting to an exponential:
Too early, or too late..."
"Three ideas worth considering wrt wages going down:
- transition management: wage insurance
- taxation: variable-rate token tax (similar to the interest rates)
- redistribution: Universal Basic Capital (UBC)" -
Danny Buerkli:
"There are only two ways of reacting to an exponential:
Too early, or too late..."
"Three ideas worth considering wrt wages going down:
- transition management: wage insurance
- taxation: variable-rate token tax (similar to the interest rates)
- redistribution: Universal Basic Capital (UBC)" -
Danny Buerkli:
"There are only two ways of reacting to an exponential:
Too early, or too late..."
"Three ideas worth considering wrt wages going down:
- transition management: wage insurance
- taxation: variable-rate token tax (similar to the interest rates)
- redistribution: Universal Basic Capital (UBC)" -
Danny Buerkli:
"There are only two ways of reacting to an exponential:
Too early, or too late..."
"Three ideas worth considering wrt wages going down:
- transition management: wage insurance
- taxation: variable-rate token tax (similar to the interest rates)
- redistribution: Universal Basic Capital (UBC)" -
Danny Buerkli:
"There are only two ways of reacting to an exponential:
Too early, or too late..."
"Three ideas worth considering wrt wages going down:
- transition management: wage insurance
- taxation: variable-rate token tax (similar to the interest rates)
- redistribution: Universal Basic Capital (UBC)" -
Our most exciting AMLD yet! The 9th edition is packed with stellar tracks and workshops.
To everyone who submitted: THANK YOU! The quality was off the charts, making our choices incredibly tough. Even if your track didn't make it this time, please join us your passion and expertise are what make AMLD special.
Can't wait to see you all there!
#AMLD #ArtificialIntelligence #MachineLearning"
https://mastodon.social/@amld/113317475774716766 -
Attending "AI Safety: the Quest to Making AI Human-Centric, Trustworthy and Safe " at the #AMLDEPFL2024. Looks extremely interesting - among the speakers in Dr. Nirupam Gupta, a colleague of mine working on stable ML; Florian Tramer, who you probably know from "Red-teaming LLMs" and "GPT2 is really good at memorizing" papers and Julia Bazinska fo Lakera AI (that you know from The Lakera Gandalf password LLM).
-
Attending "AI Safety: the Quest to Making AI Human-Centric, Trustworthy and Safe " at the #AMLDEPFL2024. Looks extremely interesting - among the speakers in Dr. Nirupam Gupta, a colleague of mine working on stable ML; Florian Tramer, who you probably know from "Red-teaming LLMs" and "GPT2 is really good at memorizing" papers and Julia Bazinska fo Lakera AI (that you know from The Lakera Gandalf password LLM).
-
Attending "AI Safety: the Quest to Making AI Human-Centric, Trustworthy and Safe " at the #AMLDEPFL2024. Looks extremely interesting - among the speakers in Dr. Nirupam Gupta, a colleague of mine working on stable ML; Florian Tramer, who you probably know from "Red-teaming LLMs" and "GPT2 is really good at memorizing" papers and Julia Bazinska fo Lakera AI (that you know from The Lakera Gandalf password LLM).
-
Attending "AI Safety: the Quest to Making AI Human-Centric, Trustworthy and Safe " at the #AMLDEPFL2024. Looks extremely interesting - among the speakers in Dr. Nirupam Gupta, a colleague of mine working on stable ML; Florian Tramer, who you probably know from "Red-teaming LLMs" and "GPT2 is really good at memorizing" papers and Julia Bazinska fo Lakera AI (that you know from The Lakera Gandalf password LLM).
-
Attending "AI Safety: the Quest to Making AI Human-Centric, Trustworthy and Safe " at the #AMLDEPFL2024. Looks extremely interesting - among the speakers in Dr. Nirupam Gupta, a colleague of mine working on stable ML; Florian Tramer, who you probably know from "Red-teaming LLMs" and "GPT2 is really good at memorizing" papers and Julia Bazinska fo Lakera AI (that you know from The Lakera Gandalf password LLM).
-
According to him the main OpenAI ChatGPT advance was UX + fine-tuning.
And that's it for AMLD Conference Generative Learning!
FIN
#AMLD #AMLDGenAI40/🧵
-
Flywheel of test-train-correct-retrain-... => Will RLHF scale (small tuning) or will keep it rolling?
Question: experimentation in high-risk area (healthcare).
Tay is getting mentionned! (fun story, back in 2015 I initially thought @SwiftOnSecurity actually was the actual Tay, because the impression was so good).
More suggestions about human-in the loop (that makes me once again want to cite one of the last @pluralistic 's articles).
39/🧵
-
Pretraining vs Fine-tuning?
OpenAI 2022 - Instruct; (but it was alredy stated in the 2018 GPT-1 paper and done by someone else back in 2020-ish).
Switch to conversational implicature.
Foundational Models fail with it, but not conversationally fine-tuned. Larger size does not help, more training date neither, nor multi-talsk ones.
(Hypothesis: destructive pruning of the neurons during fine-tuning is very powerful: maybe he should talk to Evelina Fedorenko about that)
38/🧵
-
Edward Grefenstette, director of research at Google Deep Mind.
Talk: 4 waves of computational revolution (funny that google search and Web 2.0 are not mentioned).
General formula : inference of joint probaility by optimizing a set of parameters.
=> General translators/autocompelters (T5)
=> joint code + execution (aka the compiler)So why LLMs revolution is happening now?
Nice Chinchilla excerpt.
37/🧵
-
Good robot: qualitative description to metrics to imrpove the robot description.
36/🧵
-
Generalization:
- LLMs being plugged instead of GA algoritms over the "DNA" of machines.
- LLMs for fabrication of designed forms
- Use LLMs to simulate the likely falure modes of physically simulated solutions(in my opinion still does not solve the predifinite vocabulary of DNA in robotics problem).
How to discribe a good robot?
(Evo: reproduction, but it is not acceptable for human assistants).35/🧵
-
Next up - Josie Hughes, Prof at EPFL, robotics design.
=> Embodied intelligence (#ALIFE themes - yay!)
Problem; depends on the human factor, which needs to be multi-disciplinary (I see this is an emergent space where LLMs are supposed to be a solution. I am afraid this will lead them to become a "Google university 2.0" instead...)
GPT3 was used to support iteration (Nature Machine Intelligence 2023 (why am I not surprised...) - so not yay)
34/🧵
-
(I disagree - things such as generalization in rare cases matter).
Nvidia HPARAM kicking vs examples of Paretto Curves + stacking recipes.
Sweeps of hyperparams
> mosaicML github acceleration library link.
33/🧵
-
Goal: time and $ necessary to train networks
- Same cost of inference
- Same hardware and software
- Have to optimize baselinesRecipes examples :
- Selective backprop. Drop examples that has low loss. Don't work raw, because of torch autograd, but selection with forwardpasses can work well (eg low-resolution losses)Pb:
- What is just as good?
- What is acceptable loss of good?=> Paretto frounteers
32/🧵
-
Time for Jonathan Frankle, Chief Scientist of MosaicML (acquired by Databricks).
Focus on the price of training models (BERT ~ 50$ / DeepDiffuse ~50k$ GPT3 ~ 300k$)
Chips: Nvidia ones do not drop down
Compiler => math
But Deep learning is approximate, so we can cut the costs violently
Algorithmic advances are hoped to allow us to enable to accplearte
ResNet-50 accelerated from 3h33m to 25m (7100%) on 8xA100. Team of 5 researchers over 1 year.
=> GOOD investment.
31/🧵
-
Value creation argument to push the last point forwards (I disagree, value is created by people, captured by LLMs).
EU entreprenariat failure being blamed again; but to Carmela the problem is the monopoly, not the location of the monopoly.
30/🧵
-
Talk turns to regulation & enforcement, Carmela brings example of drug regulation compared to algorithms regulation to deployment.
Next question is a re-iterations of a paperclip AGI argument about climate change (and AGI killing people to stop it). Marcel is confused, Sabine not concerned, because people kill people, not AGI.
Next up: None of the actors are european; can we regulate things we cannot make? Gaetan: good question, but yes, we can regulate: Bruxells eff.
29/🧵
-
The talk now turns to education, that pannelist agree have to change, but that is struggling (focus on Swiss politics as why it is not changing).
Next topic: is it possible to detect synthetic data in principle. Carmela mentions that it is possible that you swim in the synthetic data alrealdy. Sabine suggests that text is impossible; image varies and video remains for now possible.
Next: Does it matter if the data is synthetic. Pannel agrees that translation+ is fine.
28/🧵
-
(my take - given that it was my topic of research for the last couple of years - is best summed by a quote from Patton)
(I also disagree with Sabine that kids are more resilient than older people to disinfo. They are, but they don't believe anything anymore).
27/🧵
-
Marcel asks for their biggest worry regarding LLMs:
- Himself is worried about social ineqaulity
- Carmela is worried about uniformization of opinion driven by few, leading to less freedom
- Gaetan is worried about privacy and survelliance, specifically for policing
- Sabine becomes a devil's advocate by saying privacy does not matter, because of the flood of fake information, that does not appared fake
- Carmela is responding to say that no AI needed for disinfo#AMLD #AMLDGenAI
26/🧵 -
Carmela makes a point of the capture of the LLMs spaces by for-profit companies. If they become a tool, the providers become monopolies and start capturing massive amounts of data.
- Gaetan talks about the example of past Google dominance in search
- Gaetan suggest that there is a possibility that models maintenance can likely relegate a lot of people to poverty
- Sabine chimes in, saying she is worried about the data ownership, not the models one.25/🧵
-
Sabine chines in to mention how a gov report from April 2023 stated privacy was not needed because it does not endanger access to market in Switzerland.
Second topic: IP on Gen ML:
- IP is there to encourage creation and protect authorsSabine:
- Most politicians are lawyers, which deal with things that already happened. Precedents; how it was used, ...
- This makes looking forwards to future for them pretty darn hard
- and bridges to build with eng/sci harder24/🧵
-
First conversation:
- Devil's advocate from Gaetan: is privacy overhyped? People click through cookies, so the price is nul.
- Carmela retords with "Nobody cared when likes were sold to Cambridge Analytica, but a lot of people were ready to pay good money to have avoided that when Jan 6th 2021 came around".I would have gone with a more economical "nothing when you configure cookies, everything when that data is used to refuse you life-saving treatment".
#AMLD #AMLDGenAI23/🧵
-
Carmela Troncoso:
- Safety is "make it simple, stupid, to make it safe"'
- ML goes in the opposite directions and acheives nice results22/🧵
-
Sabine Susstrunk:
- Research on deepfakes
- Startup helping creating sketches for adverising and film industries
- President of the Swiss Science Council; advising Federal Council on the AI/LLMsGaetan de Rassenfosse:
- Economist studying the innovation
- Way AI will affect the way we invent
- Study the way the current policy framework is fit for AI
- Example:
- Combinatorics of innovation, in the context of burden of knowledge
- AI will solve that burden pb21/🧵
-
Ok, next up is the pannel on societal risks, led by @marcelsalathe, with Gaétan de Rassenfosse, Carmela Troncoso & Sabine Süsstrunk.
A refreshing change of pace:
- Marcel asking to stop with existential AGI threat risk FUD (yay)
- Marcel reminding that non-existential risk are very real and very damaging (discrimination, copyright, misinformation, ...)
- And finally reminding that the former distracts from the latter.20/🧵
-
Lack of consensus on usefullness of LLMs in the clinical context => Interesting, but I disagree. Usually it means the doctors disagree on the context that doctors are missing and disagree on.
Another nice point about replacing human tasks with ML vs new frontiers that will become available for humans augmented with ML tasks. => I really like this point of view.
19/🧵
-
Pivot to the triad of Mode/Action/Policy+Capacity around AI-guided work.
Switch away from explainability (something I like tbh).
EHR data is a "language" (but failing to update in time).
Interesting analogy with the car safety and LLMs: car is tested by NIST; humans with DMV driving test.
LLMs are currently claimed to pass the test alone based on 20 questions.
For human-augmented LLMs this is likely to change.
18/🧵
-
Next up: Nigham Shah (Stanford): Shaping and adoptions of Large Language Models in Healthcare.
Chief Data Scientist, Stanford Healthcare System.
=> Robustness, Explainability, Safety, Effectiveness and Ethicality of AI application in healthcare.
1300 different IT systems are supporting Stanford Healthcare opearitons.
Diagnosis vs Prediction. Specifically Prediction of Dealth in 12 months.
Prediction results in an advanced care conversation with patient & Med team
17/🧵
-
BERT pre-trained and fine-tuned for prediction of re-admission in a NY hospital chain (7M notes, 300 k clients).
Showing true positives of the modle prediction, but overall, missing data on false positives.
Also, perhaps unsurprisingly, no discussion of the confirmation habituation (aka doctors in the loop actually aren't - as per @pluralistic 's recent blogpost: https://pluralistic.net/2023/08/23/automation-blindness/#humans-in-the-loop).
16/🧵
-
Next speaker - Kyunghyun Cho on Health system scale language models. (PhD with Bengio, Meta AI, Genentech, ...)
Examples of what is there:
- Biopsy slides > annotation cancer locations
- Biopsy > Treatment
- Drug DesignInteresting take on healthcare as "not curing diseases" but "providing services" > EHR > Clinical an operational decisison.
Prediction of Insurance company denying claim portrayed as a difficult taks (hit - it is not; denial is automatic in the US)
15/🧵