Chester Beard

Prompt Engineering is slowly dying. Here are the 3 latest OpenAI upgrades that are making learning prompt engineering less relevant: - o1: OpenAI’s o1 has incorporated a prompting technique called Chain of Thought.  You don't need to induce the model to produce intermediate reasoning steps anymore. Now it does it itself. - DALL-E: DALL-E 3 does Image Prompting for you. It generates prompts from simple ideas written in plain English - GPT Builder: It does Role Prompting for you when creating a GPT from scratch. It also creates instructions specifying the GPT's goal and behavior. Yes, it takes time to incorporate a prompting technique into the products (the Chain of Thought paper was released in January 2022) but still learning prompt engineering is becoming less relevant for most users. This is no surprise since the goal of OpenAI is to make its products accessible and user-friendly for a broader audience. In the article below, I dive deeper into this topic 👇

13

1 Comment

🚨 Fascinating AI paper alert: "Consent and Compensation: Resolving Generative AI’s Copyright Crisis" by Frank Pasquale & Haochen Sun is a must-read for everyone interested in AI, copyright, and artists' rights. Quotes: "The opacity and scale of AI systems is disrupting the knowledge ecosystem by significantly eroding authors’ proprietary control of their works, well beyond extant digital practices that have already undermined many authors’ well-being. Whereas prior scraping at scale tended to be focused on the non-expressive aspects of works (such as facts), AI is focused by many prompts on their expressive dimensions. Search engines have historically provided links which lead users to works themselves. In contrast, AI tends to provide substitutes for such works, while failing to provide citations to the works in the dataset most similar to the texts, images, and videos it presents as a computed synthesis." (pages 8-9) - "Under the proposed mechanism, copyright owners can first request AI providers to take actions to effectively prevent their systems from generating outputs that appear identical or substantially similar to relevant copyrighted works. A copyright owner would be entitled to send a notice to an AI provider when he or she identifies that an output generated by the provider’s AI system contains either a verbatim or substantially similar copy of his or her work, or a derivative work. In the notice, the copyright owner would be obliged to document the unauthorized reproduction of the work and his or her copyright ownership, along with a digital copy or an online link to the work." (page 21) - "Given the complexity of the AI supply chain, particularly with respect to generative AI, it is not feasible to impose a per-device cost on AI providers. However, other triggers for payment are possible. Levies on the use of particular datasets may be imposed, or on model training, or on some aggregate number of responses provided to users, or on paid subscriptions. Alternatively, the level of the levy could be benchmarked with respect to some percentage of AI providers’ expenditures or revenues" (page 39) ➡ Link to the paper below. #AI #copyright #consent #AIregulation #AIpolicy #AItraining

170

25 Comments

Just dove into a thought-provoking piece on The Decoder about Voice Craft, an open-source voice cloning model that's making waves and stirring up debates over ethical concerns. This tech is pushing the boundaries of what's possible with voice synthesis, sparking a major conversation about the implications for privacy and consent. How do you feel about the balance between innovation and ethical responsibility in AI development? Could be a big topic for us to explore in our projects or discussions. Catch the full details on the debate here: https://lnkd.in/gZPFAXjg #TechSherpa #AI #Ethics #VoiceTech

3

1 Comment

What's news? Anthropic x Claude AI - my best favorite tool! New Projects feature: ✅ Available for Claude.ai Pro and Team users ✅ Allows organization of chats and knowledge in one place ✅ Includes a 200K context window (equivalent to a 500-page book) ✅ Powered by Claude 3.5 Sonnet Key benefits: ✅ Grounding Claude's outputs in internal knowledge ✅ Custom instructions for each Project ✅ Faster start and extended skills for various tasks Artifacts feature: ✅ Dedicated window for generated content (code, text, graphics, etc.) ✅ Enhanced coding capabilities with larger code window and live previews ✅ Available as a feature preview Sharing capabilities (for Team users): ✅ Share conversation snapshots in team's shared project activity feed ✅ Improves innovation and knowledge sharing across the organization Customer spotlight: North Highland ✅ Using Claude for various tasks like proposal writing and document analysis ✅ Reported 5x faster completion of content creation and analysis tasks Future developments: ✅ Continuing to improve ease of use ✅ Planning to expand project knowledge integration via native integrations with popular apps and tools This update marks a significant evolution in enterprise AI, transforming Claude from a standalone chatbot into a collaborative tool deeply integrated with professional workflows. By allowing teams to ground AI in their specific context and share insights, Anthropic is paving the way for more effective human-AI collaboration in specialized work environments. Of course, Claude continues to get a great job! My favorite tool! View all post: https://lnkd.in/gciWuHNg #Claude #AI #Anthropic #GenerativeAI

2

With the emergence of AI, the process of designing UIUX is undergoing a profound revolution. In some respects, it's a step backwards, or at least sideways, in that we are replacing complex but static, predictable interfaces with interfaces that reflect the current moment in a conversation that may not even seem like a specialized web app at all, especially once you start dealing with multimodal AI.

13

1 Comment

3-minute read (thank you SwissCognitive for sharing): Last week, OpenAI released a new model called o1 (previously referred to under the code name “Strawberry” and, before that, Q*) that blows GPT-4o out of the water in complex reasoning. Unlike previous models that are well suited for language tasks like writing and editing, OpenAI o1 is focused on multistep “reasoning,” the type of process required for advanced mathematics, coding, or other STEM-based questions. It uses a “chain of thought” technique, according to OpenAI. “It learns to recognize and correct its mistakes. It learns to break down tricky steps into simpler ones. It learns to try a different approach when the current one isn’t working,” the company wrote in a blog post on its website. #ai #openai #humanintelligence

5

I plan to implement a lot of these tools into my process.

1

Agentic design has the potential to bring the effects of HALLUCINATIONS to 0! LLMs are always going to hallucinate, so it is not about ignoring that it exists, but it is more about finding ways to mitigate those effects. When we build software involving LLMs, we want to avoid use cases where LLMs are not performing well and utilize their capabilities on tasks they are known to be good at! When we think about agentic design, we need to think about a combination of multiple LLM actors with limited agency and human domain expertise to architect the way agents interact alongside software that is going to impose a rigid structure between the different agent interactions. There are tasks where LLMs are very unlikely to hallucinate. For example, summarizing text, choosing between a limited set of options, and even writing very simple code. So, the game is about creating agents with very limited control and simple outputs such that hallucinations are as unlikely as possible and utilizing human domain expertise to rule how the different agents can interact with each other. If we need to solve complex problems, then we need to include more of those simple agents and the complexity will be offloaded on the human-designed architecture and the software needed to orchestrate the interactions. We often dismiss LLM pipelines as "Open API wrappers", but we underestimate the value of the domain expertise needed to build those wrappers to solve problems in a niche domain. If we build those agents to only accept or emit structured data such as JSON or other data structures, and considering the limited scope of each agent, it is also much easier to build software to validate the data coming in and out of those agents. For example, if an agent is only supposed to extract keywords from text in a JSON format, we can easily programmatically check the validity of the output format and the existence of those keywords in the original text. If an agent is supposed to output code, we can directly compile it and unit test it. We can also implement the reflection pattern with feedback loops where we have special agents in charge of validating the output of other agents. It is not about reducing hallucinations, but it is about building around them to mitigate those effects. In cases where hallucinations are more likely to occur or when decisions should not be taken automatically, then it is important to implement software with humans in the loop to validate the intermediary outputs. This allows humans to provide feedback to an autonomous system and brings the system back on track if it starts to diverge from what is expected. So, balancing human expertise and software with LLM agency is where the complexity lies when implementing agentic systems. There are still too many cases where it fails, but we collectively start to understand better what works and what doesn't! -- 👉 Don't forget to subscribe to my ML newsletter https://lnkd.in/g4iKyRmS --

107

19 Comments

How EmbedGPT Enhances Reader Engagement for Publishers

2

I thought that OpenAI's announcement of GPT-4o was pretty impressive, and it all boils down to the focus on users. There's still a lot of discussion about how various models will achieve better performance according to certain benchmark metrics or come closer to achieving general AI, whatever that means. It is true that the performance of an AI model is relevant to its capabilities, but let's not forget that when it comes to impacting the lives of regular people, you need technology that's easy to use and understand, and that's where the GPT-4o demo really stood out. First up, there's the desktop app, which is great to see. It would be very easy for OpenAI to say, "No, we're just going to stick with the web perpetually and that desktop app use is not worth our time." Producing a desktop app is probably not a huge lift for the company, but it also shows that the company is thinking about how people use their tech in real-world situations. Similarly, the voice assistant demos that were shown as part of the GPT-4o release have obvious implications for voice assistants of all sorts. These demos were not perfect, but they worked reasonably well and in some parts, they actually showcased capabilities that already exist but that people may not know about. For example, the voice assistant is available today in prior versions of GPT-4o, and it uses a voice suspiciously close to Scarlett Johansson, which I've seen some people remark on. This is not new (and before you write a hot take about the assistant's flirty voice, note that it's not the default voice, and one of five available). The point is that at the moment, the most capable AI models can do things that are well beyond the expectations of an average user. The issue, however, is actually getting the model to do what you want and making the model easy to interact with. Slapping a chat interface onto an AI model was a great start, but it was only a start, in the same way that developing a graphical user interface for an operating system was just a new start, not a final destination, for operating systems. There's a lot of practical user interface improvements that could be made so much so that I think even if we saw absolutely zero improvements in the capabilities of AI models, we'd still see software developers spinning up new software based on them for years, probably decades. In the long run, I do not think it will be the company with the best AI model in terms of benchmark performance that sees the most success and makes the most useful tech. As usual with technology, the company that does the best will be the one that is most successful in connecting people with its technology. What OpenAI's demos are showing to me is that it is thinking about this seriously, and that is good news for ChatGPT in the near future. #ai #llm #openai #chatgpt4o #chatgpt https://lnkd.in/gzwMc3Nm

3

A good presentation on what's becoming one of the hottest areas of discussion today: AI Agents.

6

2-minute video (thank you Dovid Schick for sharing): Google DeepMind posted this astonishing demo last month. The tester interacts with a prototype of AI agents supported by Google's Project Astra multimodal foundation model, Gemini. There are two continuous takes: one with the prototype running on a Google Pixel phone and another on a prototype glasses device. The agent takes in a constant stream of audio and video input. It can reason about its environment in real time and interact with the tester in a conversation about what it is seeing. The future is here and the implications far-reaching. #google #deepmind #innovation #ai

2

Is gen AI really just IP laundering? This is the question I asked of the USPTO following their recent open discussion on the impactvof AI. In the same way money laundering is an attempt to disconnect illicit profit from its source and associate it with a fake source to give it fake legitimacy, is AI more or less a scheme to steal IP and blur its source with the intention of pretending that the bot generated it? I had this thought while observing the USPTO open discussion on AI. This is eerily like Napster. The goal: Confuse what is the source to legitimize theft. We all know who was the real source of the stolen music then. Stolen visual and other web-based IP is harder to trace to its roots. However, ironicaly, it should become one of AIs best special skills to find likely sources of stolen content, help confirm or disprove that association and send the infringer a bill for payment due. In this open USPTO meeting, there seemed to be a lot of people speaking who are anxious to declare the entire body of prior art as in the public domain. Why? possibly so they can lay claim to it? All input is free, output is for sale. Sweet. Bottling tap water is a sweet deal. But when drawing it from someone else's fawcet from someone else's well, that's called theft. Mixing stolen water with other stolen water does not clean away the stolenness, no matter how hard you stir. One of my first impressions of gen AI is that it is another napster. It seems this is turning out to be true.

5

4 Comments

Microsoft has tweaked, then delayed its controversial AI-powered 'Recall' feature. Earlier this month, I wrote about how Microsoft's upcoming Recall feature is a train wreck in slow motion. 𝟭/ 𝗧𝗵𝗲 '𝗥𝗲𝗰𝗮𝗹𝗹' 𝘀𝘁𝗼𝗿𝘆 The service works by capturing a screenshot every few seconds. Everything onscreen is converted into text and stored to help users recall everything. The original plan was to put Recall on a bunch of new AI-centric Copilot+ laptops before eventually releasing it on Windows 11 to all. But cybersecurity experts have warned that it may have devastating consequences for users compromised by malware. - It gives hackers a single database to target. - Puts a treasure trove of historical data at risk. (Read: https://lnkd.in/gvgJveJ5) 𝟮/ 𝗜𝗺𝗽𝗿𝗼𝘃𝗲𝗱, 𝘁𝗵𝗲𝗻 𝗱𝗲𝗹𝗮𝘆𝗲𝗱 In response to the backlash, Microsoft initially announced changes to how it works and made it optional by default. Then last week, it announced that Recall will be delayed. Initial response: - Made optional by default. - Secured by Windows Hello. - All snapshots fully encrypted. Latest: - Restricted to Windows ('beta') Insider programme. 𝟯/ 𝗪𝗶𝗹𝗹 𝗯𝗲 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲 '𝘀𝗼𝗼𝗻' Microsoft cites its desire to provide a "trusted, secure and robust experience" to all customers for delaying the public release of Recall. However, Microsoft still has plans to make Recall available to all Copilot+ owners soon. How soon? I imagine even Microsoft doesn't know. But it'll be out eventually - hopefully implemented in a far more secure and robust manner. 𝗣𝗵𝗼𝘁𝗼 𝗖𝗿𝗲𝗱𝗶𝘁: Unsplash/Anwaar Ali --- My name is Paul Mah and I write #EverydayTechStories 📆 Get weekly updates: www.techstories.co/updates 👀 See my other posts: www.techstories.co 🙋 Follow me on LinkedIn: https://lnkd.in/gu5EMKQg #cybersecurity

42

16 Comments

According to a newly-released document, OpenAI is "exploring" how to "provide the ability to generate NSFW content in age-appropriate contexts." Its definition of NSFW content includes erotica, extreme gore, slurs, and profanity. my story: https://lnkd.in/gRwxTT2Z

19

2 Comments

OpenAI announced GPT-4o, its latest large language model with big improvements to text, voice and video capabilities... But those hoping for news on a search engine or rumored collaboration with Apple will have to wait.... EDIT: Updated with some spicy thoughts from Gartner's Chirag Dekate, Ph.D. about how GTP-4o stacks up with Google's AI offering. #chatgpt #ai #genai #gpt #llm

3