Roger Kibbe

Anthropic introduces Autonomous AI Agents that use a computer mouse and screen to perform complex tasks independently. Currently available for developers only through the Claude 3.5 Sonnet API.

7

Recently, Anthropic shared its perspective on California's proposed AI regulation bill, SB 1047, with the California Assembly. While recognizing the importance of safe AI development, Anthropic has raised substantial concerns about the bill's current form, which could potentially hinder innovation and harm the United States' competitive edge in AI. Anthropic's key points include shifting the focus from "pre-harm enforcement" to a more adaptable "outcome-based deterrence" model. They also advocate for regulations that focus more narrowly on frontier AI safety to avoid redundancy with federal requirements. Additionally, they suggest giving regulatory authority to the Government Operations Agency instead of creating a new state agency. Anthropic believes that these amendments will encourage innovation in risk reduction practices and incentivize companies to adopt measures that prevent catastrophic risks. In contrast, the AI Salon Policy & Civic Engagement Guild shares Anthropic's commitment to ethical AI and safe development but has broader concerns about the bills. We believe that overly burdensome regulations may discourage AI development and lead to job losses. Compliance requirements favor large companies, potentially driving smaller startups out of the market, and overregulation might cause California to fall behind other regions in AI development. Furthermore, compliance under penalty of perjury imposes significant personal risk on developers, and strict regulations could disadvantage California-based companies in the global AI market. On behalf of the over 1700+ members of the AI Salon, we urge our larger community to engage in this crucial dialogue about AI regulation. By advocating for well-balanced policies that protect public interests without stifling innovation, we can ensure California remains a leader in ethical and responsible AI development. Please sign the letters of support and join us in shaping a future where AI is developed and deployed with ethical integrity, social responsibility, and inclusive governance. Together, we can bridge the gap between AI technology and policymaking, fostering an environment where innovation thrives alongside robust safety measures. https://lnkd.in/e6MRBmcX #AIRegulation #EthicalAI #Innovation #PolicyEngagement #Anthropic #AISalon #PublicPolicy #TechPolicy Kyle Shannon Leah Fasten Brandon Tidd Stacie L. Matthew Karabinos, MAT Vikki Baptiste Jon Potter Jon Leland Cyndi Coon Daniele Hall

11

From where I sit in the SF Bay Area, cradle of innovation for at least the last 30 years, I'm a true believer in the promise of the #AI revolution. That said, I find it fascinating to consider views on the opposing side. I'm sharing below two opinion pieces from the last two weeks suggesting that the AI revolution is fundamentally overhyped and a bubble ripe for bursting, and that AI tools are actually not useful or transformative. Contrast this with venture capitalist Kai-Fu Lee's prognostication that 50% of white collar jobs will be displaced by #AI within three years, and that OpenAI and the other cutting edge movers in the space are headed to a trillion+ dollar valuation. Which side will win out? Where will we be by 2030 in the AI race? I'd love to hear your thoughts in the comments.

7

3 Comments

I do a lot of AI automation work, much of it involving batch processing and reprocessing of vast amounts of text. The low cost and high quality of Claude 3 Haiku (25 cents per million tokens in and $1.25 per million tokens out) has been fantastic. But today, OpenAI released GPT-4o-mini for use in API calls, driving the cost of AI automations even lower: 15 cents per million tokens in and 60 cents per million tokens out! To give you a sense of the savings: I once accidentally ran my batch processor using the high-end GPT-4o model. With 12 million tokens submitted and 2.5 million tokens returned, it cost me $97.50. With GPT-4o-mini, today that cost is only $3.30. That’s over a 96% savings! 😲 I have reviewed the Llama 3 8b + 70b Meta open-source models for these automations. The cost is lower still (the model is free, but hardware time is not). However, in my tests, the quality just isn't the same. Plus, considering the additional effort to host an API to access the Llama 3 models, my engineering time and money have so far been better spent using the low-cost models from OpenAI and Anthropic. Anyone else finding similar results in savings + quality?

35

1 Comment

Exciting update: I will be reviewing AI products in detail, no BS, with real world scenarios, raw, for real. 🫡 🫶 Follow to stay informed. First: @Magnific_AI Full review: https://lnkd.in/gbh5h4Wg Exceptional Quality, Slow Processing, and High Costs for Creative Image Transformation Onboarding 🚀 Pros: - Fast Sign-Up: The sign-up process is quick and requires minimal steps. - Immediate Access: After sign-up, users are directly introduced to the pricing plans, allowing for quick initiation if they choose to subscribe. Cons: - Pay-to-Play: Users must subscribe and pay upfront to begin using the tool, which can be a deterrent. - No Free Sandbox: There's no free trial or sandbox mode to test the product before committing financially. Quality of Output 🎨 Evaluations: - Accuracy: The accuracy of upscaling images is mixed. For high-quality images, the output is impressive, but for lower-quality images, the results can be unpredictable. - Creativity: The AI shows a level of creativity but can go overboard, altering critical image elements. - Overall Quality: When the criteria are met, the upscaling and image transformation are highly detailed and sharp. Speed of Generation ⏱️ Time Measurements: - Initial Uptime: Starting the process takes between 1 to 2 minutes. - Completion: Total time for upscaling is approximately 5 minutes for each high-resolution request. - Efficiency: Compared to other AI tools, #Magnificai is relatively slower. Cost of Use 💰 Pricing Models: - Pro Plan: $39/month - Premium Plan: $90/month - Business Plan: $299/month - Free Trial: 2 months, with the option to cancel at any time. - Hidden Fees: Tokens/credits used for each operation can add up, increasing overall costs. Customization and Flexibility 🔧 Assessments: - Prompt Customization: Offers some capabilities to adjust settings, including HDR intensity and creativity calibration. - Model Fine-Tuning: Limited options available and every change costs tokens and consumes additional time. - Adaptability: Relatively low as exploring different configurations is resource-intensive.e Language and Multilingual Capabilities 🌍 Proficiency: - Language Support: Limited to English with no apparent support for other languages. - Non-Standard Dialects: No evidence of capabilities in handling regional dialects or idiomatic expressions. Collaboration Features 🤝 Features: - Multi-User Access: No features supporting multi-user access or team collaboration. - Real-time Collaboration: No real-time collaboration features are evident. - Shared Projects: Lacks functionality for sharing projects within teams. User Feedback Mechanisms 🗣️ Features: - Feedback Collection: No built-in mechanisms for collecting user feedback on the platform. - Satisfaction Adjustment: No features for rating or adjusting outputs based on user feedback. Report created by Pactto's AI 🔥 Have you played with it? Thoughts?

13

Interesting article on the challenges of implementing AI regulation within existing U.S. agencies: https://lnkd.in/gv7hKq-9 It’s going to be an interesting time for any company building or leveraging AI. The need to have strong yet flexible processes in place for AI governance is becoming critical.

9

SB 1047 is a pending California law that is essentially outlawing large open source models. If you like open source (and I do!) this would be very bad. And now it is reported that one of the advocates of the bill is trying to cash in on it 🤦‍♂️ The bill technically doesn't outlaw large open source models, but it requires "PWC type audits" costing "on the order of hundreds of thousands of dollars to single digit millions." As most open source authors or university research labs don't have that much money, it achieves the same outcome. It would also make AI start-ups vastly more difficult and expensive. The AI community (or at least the part of the AI community that likes open source) unsurprisingly are trying to stop it. Today, the article below reports that "Dan Hendrycks, an executive at Center for AI Safety — the firm whose lobbying arm helped draft the bill — co-founded a company called Gray Swan that offers AI safety compliance tools that seem positioned to provide the type of auditing data the bill would require". https://lnkd.in/gdutbeKq

64

14 Comments

Whether to include LLM-based GenAI features in your data product is not a simple, binary question. Strong leadership is needed, and context really matters—particularly when we have this risk of reacting to incorrect information that looks plausible and believable because of how many LLMs tend to respond to prompts with a positive sheen much of the time. Learn more below. ⬆️ Join 2k readers: tap “Visit my website” to get my latest thinking in your inbox 🚫 Don’t like LinkedIn’s algo deciding what you see? Tap the bell under my profile banner to get all my posts ♻️ Please repost if this was useful 🎧: https://bit.ly/3W5IddY #LLMs #genAI #ai #ux #design

2

5 Comments

Very interested to see this unfold. The concept that someone can directly control the algorithms that shape their experiences is not new and maybe inevitable. It would require companies who deliver algo-powered experiences to relinquish control and accept the reality that maybe they don't know what's best for the user. Instead of trying to tune the algorithm/LLM using signals gathered from UI and content interactions to predict what a user wants (which also means trying to understand and predict intent behind behaviors 🤯 ) we instead focus on building tools to give direct access to the algorithmic systems that work on their behalf. Basically letting people create and shape their own data model. 🤔🤔

18

2 Comments

It's great to see Governor Newsom and his team appreciating the nuances of regulating Generative AI and veto-ing S.B. 1047. We need AI regulation for sure, but it needs to be grounded in how GenAI models and applications actually work. I'm looking forward to seeing more efforts in (1) better understanding of what safety means for GenAI and what guardrails effective guardrails even look like, (2) funding to academic research, think tanks, and labs to address (1); and (3) more partnerships between government, academia and industry on AI. https://lnkd.in/gVc7_7va #generativeai #airegulation #sb1047

18

1 Comment

"In a memo from 2011 titled “Antitrust Basics for Search Team,” the company recommended avoiding “metaphors involving wars or sports, winning or losing,” and rejecting references to “markets,” “market share” or “dominance.”" https://lnkd.in/gv4atzNs #Google #AntiTrust

2

Made some further optimizations to the shdn.io stylesheet so it's now less than 10kb compressed while still supporting both light and dark mode: https://lnkd.in/g3rz2V_Y

45

inception llms: experimenting with a new LLM design pattern where I prompt Gemini to write code on my behalf which, in turn, uses Gemini to perform tasks - So meta, isn't it? One prompt for one task. in less than two hours created a full pipeline which would have taken days otherwise. Gemini wrote more than 97% of my code, discoverying libraries I was not aware of - overall, saving me tons of time as swe. Where do you think you can use this power in your product? try it yourself next time you develop a feature. I named it "inception llms"

76

10 Comments

Apple released its technical report today on its new foundation models. Key takeaways: (1) Apple spends an order of magnitude less to train its models by using cheaper and fewer Google TPU v4s and v5s rather than Nvidia GPUs; (2) Apple heavily used quantization to reduce training cost but made up for the degradation in quality by changing the base generative algorithms to use Transformers with a technique called Adapters (everyone could do this with GPUs too); and (3) Apple has both on-device and in-cloud versions, allowing easily tiered fidelity responses. As a minor takeaway, Apple's on device ANE is finally used by its models. It's amazing how much power (6 watts) the ANE can actually use when most of us have seen it using near zero for other foundation models.

4

Anthropic has just introduced „computer use” for Claude 3.5. It can use the #gui like human - move the mouse, fill the forms. It looks like a killer app for me (and #RPA killer in the future when it will evolve). Claude looks like a actual winner in the AI/LLM race. It’s more stable, more secure and more usable for business cases than GPT. And it’s „less drama” (as they pointed out in their ad). #AI #LLM https://lnkd.in/d6wciU2C

16

5 Comments

‘Algorithmic programming comprises three steps: (1) Pooling: Creating a pool of eligible stories for the specific module; (2) Ranking: Sorting stories by a ranking mechanism; and (3) Finishing: Applying editorial guardrails and business rules to ensure the final output of stories meets our standards. Editorial judgment is incorporated into all of these steps, in different ways.’

1

2 Comments

7 RECOMMENDATIONS FOR HARDWARE STARTUPS It’s hard to understate the difficulty of the v1 launch of a consumer hardware company. I see a ton of new HW startups coming out. Here are some counterintuitive recommendations… 1) Target a small market for v1, not a large market. You need a small number of people to love your product and this will earn you the right to build a v2. You have to understand your v1 will not make you a successful company. Your v1 needs to prove that you deserve to exist in the first place. In the case of WHOOP, we started with pro athletes who were willing to put up with our v1’s short battery life and flaky Bluetooth connections because they so desperately needed Sleep and Recovery data. A broad consumer base would have rejected it. 2) Be great at no more than 1-3 things. If you try to be great at more than that, you will build a hardware that is mediocre at a wide range of things. You won’t think it’s going to be mediocre, but it will be. 3) You want a small launch. Not because you want small sales but because you want to have a steady ramp. It’s much better to do 100k of sales steadily over 12 months than in 2 months with a huge drop off. Keep in mind you need to manage your manufacturing partner and that there’s nothing better or more efficient than steady and predictable. 4) CAC is your enemy. Referrals are gold. The magic of referrals is that they signal high NPS and they help you grow steadily. Again everything that helps you have a steady HW manufacturing plan is good. Building a referral program for v1 launch is a good use of your time. 5) Be excellent at Software. Turns out being good at both Hardware and Software is a hard thing and something that your BIG COMPANY competitor is actually not great at (excluding Apple here). For some reason, there’s such a magnifying glass on the Hardware development, that Software is then overlooked by consumer startups. Don’t make that mistake. You can be great at both. 6) You should spend an unreasonable amount of time and money on ONE aspect of your Hardware. Could be the design or a technical breakthrough. But something needs to be SPECIAL. In our case, we designed a modular battery pack that allowed Whoop to be charged without taking it off your body. This made our development timeline much longer, but it allowed us to achieve something that no other wearable has achieved since: 24/7 health data. 7) Finally, have an unrealistic belief in yourself and your team. You are not supposed to succeed and the odds are stacked against you. Building hardware takes confidence and you’re up against big players who will rip you off. We went so far as to leave a message for our competitors on our circuit board (pictured below). In the case of Amazon and others, the prophecy came true. If you’re building consumer hardware, I wish you nothing but success. It will be super hard, but super rewarding. Good luck 👊🏼 #consumer #hardware #wearable #startup #whoop

4,788

214 Comments