Dennis Crouch’s Post

OpenAI's response to The New York Times lawsuit is hot off the presses. OpenAI cuts quickly to the chase, arguing that: 1. NYT used OpenAI in an improper (ToS-violating) manner to get those infringing results. For example, NYT fed into OpenAI lots of its own articles in order to have the system output other text from those articles. 2. The articles cited by NYT as having been infringed were >3 yrs old so statute of limitations applies. The funny thing about argument # 2 is that NYT had to focus on older articles because OpenAI is not trained on the latest articles. And, the longer an article has been out there, the more likely it is that the article was parroted or paraphrased elsewhere as well, which means OpenAI’s training could encompass multiple copies of the same article. OpenAI also pooh-poohs NYT’s CMI deletion claims by citing the failure of NYT to show: (i) what CMI was removed, and (ii) OpenAI’s scienter of such removal - which would help prove the CMI removal was "by design."

We saw a couple of new OpenAI patents in yesterday's livestream that showcased an interesting mix of "user interface" patents and "under the hood" patents. Here’s the key takeaway: Two patents focused on how the user interacts with the chatbot. Their claims describe the back-and-forth interaction between the user and the language model. The beauty of these patents? Their detectability in a product is clear, making them easier to enforce and protect. The other side of the coin is "under the hood" patents, which tackle detailed technical solutions happening inside the software. While these may be harder to detect, they provide deep technical innovation that strengthens the portfolio’s foundation. Takeaway: When drafting software patents, always consider shifting your perspective. Sometimes subtle tweaks in claim wording can frame the invention from the user interaction side rather than just the backend, adding both clarity and value. 💡 Does your patent portfolio strike a good balance between user-facing and backend technical innovations? Share your thoughts in the comments!

A clean (devoid of unnecessary abstractions) experimental package released by OpenAI for multiagent systems. There is no official support as of now to the package, but few interesting examples. Also like their simple take on defining agents - An Agent simply encapsulates a set of instructions with a set of functions (plus some additional settings below), and has the capability to hand off execution to another Agent. MIT license. Good read for exploration. Link in comments. [openAI, agents, LLMs]

I don't see any use of OpenAI o1 for companies... Why would they use a product that hide the CoT traces when it gives explainability and trust in the process..? Moreover the CoT traces are critical to enhance any agentic system... specially one that can be programmed like us. Any company needs data to specialize their future agentic systems. Remember agentic systems are still R&D and if you give the CoT traces to another company you are basically giving up your IP and the future outcomes of that technology.! basically killing your business. It is maybe time for us to do another workshop on that matter.. What do you think? Tell me in the comment if there is something you would love to discuss in this workshop.

Alibaba releases an ‘open’ challenger to OpenAI’s o1 reasoning model A new so-called “reasoning” AI model, QwQ-32B-Preview, has arrived on the scene. It’s one of the few to rival OpenAI’s o1, and it’s the first available to download under a permissive license. Credit : @techcrunch

"We're gonna steamroll you" - Sam Altman. Some good insights into where OpenAI is headed in this interview and a warning to those just building wrappers and broad solutions. Also a good reminder that things are going to get better quickly with these models. What doesn't work now might work in a few months and what works pretty good now might work really well soon... 🚀🤘 Most importantly, the key is to remember that this in many ways is just another cool development tool for folks designing solutions to solve real problems for real workflows. I was a programmer when XML came out and I remember the excitement and people wanting to use it for everything. It is still around in many solutions (including heavily in the background in Microsoft Word) but you never hear about it, stuff just works. For those law firms and vendors building: 🤼 Figure out what new use cases are possible to be solved now that couldn't be addressed in the past 💫 Make the user experience is simple with a great UI/UX so it looks like magic 🧩 Add value by layering in business knowledge to the tech (no prompt engineering for users) ✅ Make sure results are consistent and accurate 😎 Assume the model is going to get better and your app will just work better when the models get better and cheaper You wouldn't just release an Access database as a solution when databases became widely available and here you can't just wrap the APIs and expect long lasting change. Just like legal tech that has been around for decades (see Doc Automation), the key is a seamless user experience and understanding existing processes to address change management challenges. The options for viable new solutions are limitless but just like all other prior apps it is going to take hard work to build quality solutions that work to solve real problems. #innovation #legaltech #openai #changemanagement #lawfirms

Very cool paper, and very relevant to OpenAI's release of o1. Use LLMs and consorts where they shine, that is, to generate plausible solutions, and then use a domain-specific verifier to pick the best solution: https://lnkd.in/gzT2htQm This is essentially shifting the burden of computation from training to inference, which commensurates nicely with the business model of most parties involved. See also this article for more details about OpenAI Strawberry and scaling laws: https://lnkd.in/gUTT2CGV. #llm #strawberry #o1

By the way, I think the OpenAI o1 model is a good example of how some parts of prompt engineering for a wide audience become unnecessary. It already integrates a chain of thought—one of the best prompt techniques, though I believe there are even better ones. The model essentially writes a prompt for you publicly before providing an answer. That doesn't mean all of prompt engineering disappears. You still need to provide enough context—who, where, why, and how—if you want more focused results. However, the "black magic" aspect of prompting gradually fades as models improve and incorporate self-prompting.

An update to an OpenAI article introduces the DALL·E editor interface, allowing users to edit images. This feature works by selecting a part of the image and then making modifications through a chat-based description of the desired changes.

OpenAI has unveiled its Prompt Engineering guide, a comprehensive resource for those looking to master the art of extracting optimal results from large language models. Click the link and start exploring example prompts! https://lnkd.in/dcihvvTj