Oren Etzioni’s Post

Harmonious.ai's weekly paper roundup is back, with a big help from AI. To streamline my reviewing, I wrote a tool that pulls weekly papers from Hugging Face, sorts them using vote count, and copies to a Google sheet. This helps me focus on just reading papers and taking notes. I then wrote another tool that turns my raw notes into more readable reviews, as well as generates an overview of those papers, all with the help of GPT-4o. Final processing step is to generate a Markdown file that can be uploaded to Harmonious.ai. The tool is on GitHub at https://lnkd.in/gGzSwbqP. With these toys, I aim to spend no more than three hours per week on reviewing and posting the reviews to Harmonious. Overview These papers collectively delve into advancements in large language models (LLMs) and their applications across various domains. Common themes include enhancing LLM capabilities for specific tasks, such as ultra-long text generation (e.g., LongWriter), and developing specialized models tailored for fields like healthcare and chemistry (e.g., Med42-v2, ChemVLM). Many works focus on improving model efficiency and usability, such as ControlNeXt for controllable image and video generation and Generative Photomontage for fine-tuned image creation. The integration of multimodal data is also a significant aspect, particularly in models like VITA and mPLUG-Owl3, highlighting the importance of vision and language understanding. Additionally, the use of self-play, reinforcement learning, and tree search methods to enhance reasoning and problem-solving abilities in models such as rStar and DeepSeek-Prover-V1.5 is emphasized, showcasing an enriched exploration into AI-driven scientific research and machine learning improvements. Spotlight: Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2 By: Google DeepMind This paper introduces Gemma Scope, a suite of sparse autoencoders aimed at making neural network interpretations more accessible. I appreciate how the authors provide pretrained models and evaluation metrics, which significantly lower the entry barriers for researchers. The interactive demo and tutorial are particularly valuable, offering hands-on experience to new users. Additionally, this work stands out as a noteworthy follow-up to the Golden Gate Claude project from Anthropic. Overall, the contributions from Google Deepmind offer practical tools for advancing safety and interpretability research in AI. https://lnkd.in/gFXMhSi8

Credit to PayPal for moving quickly on this. On August 7th they introduced a bug into their Android app that caused a spike in negative feedback from customers who were running into an infinite 'Captcha' loop while trying to login to the app. Even though they're not a paying customer (yet...), we sent them a ping to let them know that: - Our system was seeing a significant change in feedback from users stuck in the dreaded 'Captcha' flow - 98% of the feedback was from a particular user cohort (Android users) - 100% of the Android feedback was from users running App Version 8.67.0 or 8.68.0 A few days later they resolved the issue and feedback related to this problem reverted to previous levels. Now to really close the CX loop PayPal should respond to all the users who complained, letting them know this problem is fixed. Our research shows you'll see about a 25% increase in app store rating if you respond to users and let them know that 1) you're listening and 2) you've fixed the problem. Jackson Hoagland, Max Dekle, and the rest of the incredible Unwrap.ai NLP team - nice work developing a data pipeline that can accurately categorize and detect extremely specific problems among unstructured data sets.

Watching Microsoft’s Chief Questions Officer, Trevor Noah, take the ‘Real or Not’ quiz makes me feel a bit better about my own score. What a great episode of #ThePrompt which shows how hard deepfakes can be to detect, available AI tools to help check content authenticity, and what each of us can do to stop their spread on social media.   Take the Real or Not Quiz here: https://lnkd.in/exRSSkUJ Learn more about what Microsoft is doing to help protect global elections here: https://lnkd.in/eQijnf4j

Chomsky argued that LLMs learned impossible languages as well as possible ones & thus couldn’t tell us useful things about language. Nope: “Our core finding is that GPT-2 struggles to learn impossible languages when compared to English as a control, challenging the core claim.” Paper: https://lnkd.in/ePQ9etZg

While I'm optimistic about AI being used for good, there are, and will continue to be, bad actors. That's all the more reason I'm excited about the AI tools TrueMedia.org founder Oren Etzioni and his team are developing to detect political deepfakes and ensure that our elections are free and fair.

Thanks to all who engage with my last post about recruiting engineers at seed-stage startups. I'll summarize the key points and follow up in another post with one specific option: offshoring. Shashank Chiranewala wrote that founders who recruit from their networks can have a good chance. It's best to work with someone you know and trust from previous roles, and the draw of working on something new and exciting with the folks you like and respect is huge. I agree 100%: this should be your first venue. Stefan Mai suggested be more generous with equity, and I agree: it's worth paying an extra percentage or two if the next John Carmack shows up. The delicate balance is that founders take huge risks getting from nothing to funding and appropriate incentives must exist. Xu Ning and Lucas Adams discussed FAANG engineers, with Xu suggesting they may not be a fit, and Lucas saying that those who are entrepreneurial may still be good fits and it takes many tries to land one. I agree with Stefan that FAANG have amazing engineers. A subset of them are amazing at a narrow area (e.g. planet-scale, milliseconds latency data access) and may need ramping up to wearing dozens of startup engineering hats. I believe that big tech engineers who have a growth mindset can succeed at startup life. The savings they accumulated at FAANG, and a desire for a change in scenery (i.e. burn out) will be helpful. There are also stellar engineers who never worked at FAANG. They sometimes have non-conventional background (e.g. Bachelor in English), value flexibility (e.g. working from a remote cabin in Idaho), and usually have a disdain for LeetCode. I always try to pattern match this profile if possible. They can be amazing startup engineers! One of Lexion's co-founders is such an engineer (hi JamesB). Next up, I'll discuss offshoring. Marc Powell Mounish Sunkara Pablo Velasquez

Excited to finally be able to share our case study with Google Cloud! Earlier this year, Google Cloud used Yoodli to rollout first of its kind AI powered sales training to over 15,000 GTM sellers (with 92% CSAT). Context: Google Cloud Consulting needed to roll out a brand new GTM strategy across their entire org (15k+ employees). The problem: 1) Scale: Instructor led training or manager roleplays couldn’t scale to 15k employees 2) Consistency: Managing consistency of feedback was too challenging  Timing: Getting the entire org to complete training in 1 quarter was challenging Yoodli’s impact: 1) 92% CSAT (20%+ higher than typical trainings) 2) Core metric (seller hitting talking points) increased by 100% from the first to the last Yoodli practice 3) Google Cloud is now rolling out Yoodli AI roleplays across all their enablement programs - from manager training to sales coaching, pitch certification and more Read more in our case study here: https://lnkd.in/gdjFWcS9 Big thank you to Lee Moore Arpit Vaidya, Reaves Wimbish, Jason Nager, Linda Moss, Kinga Kaiser, Nathan Elequin, Ayla Lewis, Ultan Coyle, Patrick Austin and the entire Google Cloud GTM team for being awesome partners! Looking forward to scaling our partnership with Google Cloud - more updates coming soon :) If you're part of the Cloud GTM org, please reach out! Would love to hear how we can improve your Yoodli experience!