Sign in

Marco Herzog

@marcoher.bsky.social
1.1K followers 423 following 167 posts

Connecting AI research & Applied AI since 2018 | Previously: ML, NLP, RecSys | Now: GenAI, LLMs, RAG | Natural science, tech and climbing is my thing.

PostsRepliesMedia
Marco Herzog @marcoher.bsky.social · 08/12/2024
Yup, the reasoning Dialoge is derailing the model sometimes. QwQ for instance is suddenly switching to Chinese sometimes. Instruction following is a bit more tricky than for other models.
040
Marco Herzog @marcoher.bsky.social · 07/12/2024
I’ve tested the new llama 3.3 70b 🦙 and the instruction following so far is magnifique. 👌
040
Marco Herzog @marcoher.bsky.social · 06/12/2024
I’ve seen the benefits in software development, but not the downside, though. I assume because R&D is more purpose driven in form of the design and investigative work, which now shifts into a human as a judge role. aidantr.github.io/files/AI_inn...
aidantr.github.io
020
Marco Herzog @marcoher.bsky.social · 06/12/2024
What does this mean for your business? AI offers significant benefits but also presents challenges that require careful management. Leaders must ensure a positive impact on both innovation and workforce satisfaction.
120
Marco Herzog @marcoher.bsky.social · 06/12/2024
𝟯. 𝗪𝗼𝗿𝗸𝗳𝗼𝗿𝗰𝗲 𝗗𝘆𝗻𝗮𝗺𝗶𝗰𝘀: ☑ Addressing Morale Issues: → 82% reported reduced job satisfaction ↳ Due to decreased creativity and skill underutilization ☑ Reskilling Imperative: → 71% of researchers planning to reskill ↳ Reskilling in AI as a tool usage ↳ And in evaluating AI-generated outputs
100
Marco Herzog @marcoher.bsky.social · 06/12/2024
𝟮. 𝗩𝗮𝗿𝗶𝗲𝗱 𝗜𝗺𝗽𝗮𝗰𝘁: ☑ Top Performers vs. Others → Top-performing scientists nearly doubled their output. ↳ Bottom third saw no gains ↳ But importantly, no harm was done. → Top performers use domain knowledge to filter AI suggestions. ↳ Resulted in the 90:10 performance inequality.
100
Marco Herzog @marcoher.bsky.social · 06/12/2024
𝟭. 𝗔𝗜 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗶𝗻𝗴 ☑ Boosted overall performance: → AI-assisted researchers discovered 44% more materials → 39% increase in patent filings → 17% rise in product innovation ↳ AI uplifts productivity ☑ Novel Discoveries: → AI-generated compounds were more novel ↳ AI enhances creativity and innovation
100
Marco Herzog @marcoher.bsky.social · 06/12/2024
If you are a leader, you shouldn't ignore this paper from MIT for your business and your workforce. This research from a large U.S. firm's R&D department shows real-world effects of AI — positive and negative. These findings can be adopted by other businesses and departments as well. 🧵 #AI #ML
140
Marco Herzog @marcoher.bsky.social · 06/12/2024
𝟯. 𝗪𝗼𝗿𝗸𝗳𝗼𝗿𝗰𝗲 𝗗𝘆𝗻𝗮𝗺𝗶𝗰𝘀: ☑ Addressing Morale Issues: → 82% reported reduced job satisfaction ↳ Due to decreased creativity and skill underutilization ☑ Reskilling Imperative: → 71% of researchers planning to reskill ↳ Reskilling in AI as a tool usage ↳ And in evaluating AI-generated outputs
000
Marco Herzog @marcoher.bsky.social · 06/12/2024
𝟮. 𝗩𝗮𝗿𝗶𝗲𝗱 𝗜𝗺𝗽𝗮𝗰𝘁: ☑ Top Performers vs. Others → Top-performing scientists nearly doubled their output. ↳ Bottom third saw no gains ↳ But importantly, no harm was done. → Top performers use domain knowledge to filter AI suggestions. ↳ Resulted in the 90:10 performance inequality.
100
Marco Herzog @marcoher.bsky.social · 06/12/2024
☑ Boosted overall performance: → AI-assisted researchers discovered 44% more materials → 39% increase in patent filings → 17% rise in product innovation ↳ AI uplifts productivity and innovation ☑ Novel Discoveries: → AI-generated compounds were more novel ↳ AI enhances creativity and innovation
100
Marco Herzog @marcoher.bsky.social · 04/12/2024
Oh Yhea, I agree. I think the tricky part is that a really big user base, I would argue the majority of users, will be totally not aware of it for a longer time.
100
Marco Herzog @marcoher.bsky.social · 04/12/2024
Or simply prompt to vid:
010
Marco Herzog @marcoher.bsky.social · 04/12/2024
It has Voice Control Magic:
110
Marco Herzog @marcoher.bsky.social · 04/12/2024
It does Native Camera Cuts, Smooth Special Effects:
110
Marco Herzog @marcoher.bsky.social · 04/12/2024
It does image2video with face expressions:
110
Marco Herzog @marcoher.bsky.social · 04/12/2024
It can do image2video with movements:
110
Marco Herzog @marcoher.bsky.social · 04/12/2024
Tencent open sourced this, Impressive! But I think we need to start talking about AI generated content detection on social media more seriously … This open source is 🤯 More content with other capabilities in 🧵 aivideo.hunyuan.tencent.com
250
Marco Herzog @marcoher.bsky.social · 04/12/2024
This essay from DeepMind is something you shouldn't miss, it made me very excited about the future of science. A new golden age of discovery Seizing the AI for Science opportunity deepmind.google/public-polic... #AI #ML #Science
deepmind.google
A new golden age of discovery
In this essay, we take a tour of how AI is transforming scientific disciplines from genomics to computer science to weather forecasting. Some scientists are training their own AI models, while...
050
Marco Herzog @marcoher.bsky.social · 03/12/2024
If this is not telling an interesting story, than what does? Accenture makes more money with GenAI than OpenAI. GenAI revenue: OpenAI: $3.4bn Accenture: $3.7bn I’m still thinking what that means…
050
Marco Herzog @marcoher.bsky.social · 03/12/2024
Unveiling the true inspiration behind the ‘attention’ operator in Transformers! From Bahdanau’s emails to Karpathy, we learn how attention transformed neural networks. Surprising that ‘Attention is All You Need’ outshines its predecessor by Bahdanau et al. #AI #ML #Transformers
031
Marco Herzog @marcoher.bsky.social · 01/12/2024
Actually it would be possible with a browser plugin. It could track the views, clicks on the post, engagement etc. and filter the discovery feed based on the historical data.
020
Marco Herzog @marcoher.bsky.social · 01/12/2024
Yeah, I agree. It’s supposed to do that when you click the choose “Show more” or “Show less” of it, plus likes. However, it seems to be very slow or not working properly. I was considering making one myself, but it’s not possible because I don’t have access to stats about impression rates.
110
Marco Herzog @marcoher.bsky.social · 01/12/2024
You can pin it. I’m using it very often. It has positive and negative filters.
010
Marco Herzog @marcoher.bsky.social · 01/12/2024
I created a feed that works quite well. Maybe you like it. bsky.app/profile/marc...
110
Marco Herzog @marcoher.bsky.social · 01/12/2024
Super nice, it's going to be great to have you in the list soon.
110
Marco Herzog @marcoher.bsky.social · 01/12/2024
Hey Walid, I just checked your reply history which mostly consists of starter pack requests. Feel free to come back when your content is primary about AI and your interacting with the AI community.
100
Marco Herzog @marcoher.bsky.social · 01/12/2024
So many questions. Why leaving threads? Why don’t uncringe your LinkedIn, and did you mean Substack?
000
Marco Herzog @marcoher.bsky.social · 30/11/2024
😁 But I was seeing your content without the dog. I needed to check your profile to understand what you meant. One big benefit of Bluesky is creating and following custom feeds.
010
Reposted by Marco Herzog
Sujay Shukla @sujayxaradhya.bsky.social · 28/11/2024
Cursor vs Windsurf I just spent a week deep-diving into Cursor and Windsurf, so you don’t have to. Here's everything you need to know 👇🧵
141
Marco Herzog @marcoher.bsky.social · 29/11/2024
My pleasure!
000
Marco Herzog @marcoher.bsky.social · 29/11/2024
So basically majority voting based on o1? Similar: arxiv.org/html/2402.05...
arxiv.org
More Agents Is All You Need
110
Marco Herzog @marcoher.bsky.social · 29/11/2024
Good old pen and paper 👌
010
Marco Herzog @marcoher.bsky.social · 29/11/2024
Play with it here: bbycroft.net/llm
bbycroft.net
LLM Visualization
A 3D animated visualization of an LLM with a walkthrough.
230
Marco Herzog @marcoher.bsky.social · 29/11/2024
By far the best visualization and walkthrough of a Transformer I've ever seen. You can explore the algorithm down to every add & multiply, seeing the whole process in action. With animation and explanation. That was definitely created with a lot of passion and work. By Brendan Bycroft. #AI #LLM #LM
1114
Marco Herzog @marcoher.bsky.social · 29/11/2024
I think the single most important. Windsurf hallucinates less.
020
Marco Herzog @marcoher.bsky.social · 28/11/2024
A new update on the AI feed. The feed now has 66 unique negative filters. They remove any form of AI art. There are 186 filters that are filtering content related to the sphere of AI. bsky.app/profile/did:... #AI #LLM #ML
222
Marco Herzog @marcoher.bsky.social · 28/11/2024
This is very true. Would you agree that 12 likes, 6 comments and 2 reposts on 100 impressions shows you as a content creator you're doing a better job than 12 likes, the same 6 comments and 2 reposts on 12k impressions? Because I'm not able to tell currently. My content could be great or awful.
100
Marco Herzog @marcoher.bsky.social · 28/11/2024
I agree, that's why meant not on the post itself, but as stats purely for the post creator. Otherwise, I can't really tell if my content is actually creating value and gives me no way to improve my content. I think the issue you are describing will happen anyway by just seeing likes. What you think?
100
Marco Herzog @marcoher.bsky.social · 28/11/2024
@bsky.app Impression stats are really missing. Not necessarily visible for everyone on the post, but at least for the post creator. 1. To know how big the audience really is 2. To see the actual quality of post (engagement/Impressions)
110
Marco Herzog @marcoher.bsky.social · 28/11/2024
github.com/Alec2435/ama...
github.com
GitHub - Alec2435/amazon-fresh-server: An amazon fresh mcp server
An amazon fresh mcp server. Contribute to Alec2435/amazon-fresh-server development by creating an account on GitHub.
000
Marco Herzog @marcoher.bsky.social · 28/11/2024
Do you want to see the new Anthropic Model Context Protocol in action? It was announced just 3 d ago, and we already see great apps. This time build by Alec Velikanov. Take this project as an example to build something yourself. I'm off, need to cook 7 courses with the stuff Claude ordered. #AI
200
Marco Herzog @marcoher.bsky.social · 27/11/2024
What can we do about the benchmark fatigue for #LLM? More people I speak to don’t take them seriously anymore, and I can’t blame them. I still hesitate, but I’m about to drop them as well. I think we need something new for #AI eval. Or is there something I’m not aware of?
150
Marco Herzog @marcoher.bsky.social · 27/11/2024
Hey Niklas, great to see your comment. Keep on going with the good work and please come back as soon as you account is a bit more developed. I appreciate your interest and feel free to follow the folks in the list.
100
Marco Herzog @marcoher.bsky.social · 27/11/2024
What a name 😆
000
Marco Herzog @marcoher.bsky.social · 27/11/2024
The funny part is just that with 275 million MAUs it’s definitely not starting.
030
Marco Herzog @marcoher.bsky.social · 27/11/2024
Hey, I’m currently evaluating which eval lib to use. And I’ve seen just e2e pipeline testing for RAG or something like Promptfoo. We need to test fine tuned qlora models that we might even prune further and need to eval them after. So sure, let’s connect!
010
Marco Herzog @marcoher.bsky.social · 27/11/2024
I see where you’re coming from, and I agree to 70%. But with the right workflow you can get much better results. I would say a total newbie can become an intermediate dev.
000
Marco Herzog @marcoher.bsky.social · 27/11/2024
And great source of getting raw opinions and experience.
010
Marco Herzog @marcoher.bsky.social · 27/11/2024
Will check it out. What do you think about this here? Not a model at all, but maybe an inspiration. bsky.app/profile/maca...
010