Stan Podolski

One question, which might become relevant for the current "Death of Software" debate, is to what lengths OpenAI and other foundational model providers will go to obtain more training data once they've scraped and used most publicly available data on the Internet (which is likely the case soon, if it isn't already). Deals with large book publishers and movie right owners seem like a logical next step. How about internal documents from enterprises? Or large amounts of data from e.g. hospitals, insurances, law firms, or accounting firms to improve model performance in those areas? The assumption is, of course, that personally identifiable data will be removed before training. If it turns out to be too hard to buy this kind of data – and if it's important for e.g. OpenAI to get better with healthcare – they could simply buy a few hospitals. Curious to hear if people think of this is a realistic scenario.

45

23 Comments

So Delaware wants to be a “tech hub” 🏙️, but here’s the reality check: according to Code.org's 2024 State of Computer Science Education Report, only 61% of high schools here offer computer science 🖥️, and less than 6% of students are actually signing up. Meanwhile, neighboring states like Maryland and New Jersey are way ahead—fully funded 💸, with state plans 📜, and higher enrollment numbers. And what about the younger kids? We’re handing out Chromebooks in elementary school 📲, but where’s the computer science education to go with them? Kids are using tech every day but aren’t getting the skills they need to understand it, build with it, or innovate for the future. 🌐 If we’re serious about preparing our kids—and Delaware—for a tech-driven future 🚀, it’s time to go beyond talk and invest in computer science at every grade level. Let’s put some action (and funding!) 💥 behind the “tech hub” dream. #Delaware #CSforAll #TechHubGoals #FutureReadyOrNot #RealTalk

14

5 Comments

o1 Pro vs. Claude Sonnet 3.5: GTM Use Cases and is it worth it? The image below has been going around and I saw it in the original AI group on Reddit, Inc., so I thought I’d give my .02 cents on this, how it applies to GTM and instead of a prompt for you all, I’ll talk about the difference between prompting both models and my verdict for most every day users. (Image is not mine ps) We have: 1. Complex Reasoning o1 Pro delivers slightly higher accuracy (5-10%) in complex tasks but takes 20-30 seconds longer to respond where Claude is instantaneous. GTM Application: Claude helps GTM teams act quickly, while o1 Pro is better suited for deep-dive strategy sessions. 2. Code Generation Claude Sonnet creates clean, maintainable code with minimal complexity. o1 Pro tends to overengineer, (like a smart person making things way harder than it should be). GTM Application: For building automations or integrating tools like Salesforce or HubSpot, Claude shines. 3. Advanced Mathematics o1 Pro is unmatched for PhD-level calculations but overkill for everyday needs. Claude handles 95% of math tasks quickly and accurately. GTM Application: GTM leaders can rely on Claude for tasks like revenue forecasting or quota optimization. 4. Scientific Reasoning o1 Pro excels in depth and nuance. Claude provides clear, concise responses, better suited for quick communication. GTM Application: GTM professionals need clarity for business cases or stakeholder updates. Now how do you prompt between the 2? Key Differences in Prompting: o1 Pro Prompting • Prefers extremely simple, direct prompts • Works best with minimal instructions • Avoid step-by-step guidance • Basic delimiters (###, XML) are sufficient • Less is more - shorter prompts work better • Doesn't need chain-of-thought prompting Claude Sonnet 3.5 Prompting (and applies to ChatGPT 4o) • Benefits from detailed instructions • Handles structured XML-style formatting well • Works effectively with role-based prompts • Can process extensive context and documentation • Responds well to specific formatting requests • Benefits from clear deliverable expectations General Prompting Differences: • o1 Pro needs minimal intervention while Claude accepts detailed guidance • o1 Pro works better with implicit context; Claude handles explicit context well • o1 Pro may get confused with too much instruction; Claude remains stable • o1 Pro prefers letting its reasoning take over; Claude follows structured guidance • Claude handles longer, more detailed prompts without performance degradation • o1 Pro performs better when allowed to "think" independently Verdict so far: for 97.42% of people, my recommendation is staying with the $20 Claude or ChatGPT subscription. It is nice having a larger context window and better abilities overall with the $200 option, but not sure it’s worth $2400 a year when other options like Claude are nearly as good. I’ll keep testing, may change my mind ;)

411

34 Comments

Skill but not Effort Drive GPT Overperformance over Humans in Cognitive Reframing of Negative Scenarios https://lnkd.in/e-Bc4BbA

4

There are hundreds of AI models out there 🚀 Every vendor claims theirs is the best 💪 How do we know who to trust? 👉 Benchmarks! But, AI benchmarks are like the Wild West today. In this interview, Dylan Fox, Founder/CEO of AssemblyAI and I do a deep dive into what AssemblyAI is up to and how the industry should do benchmarks. Here’s what stood out to me most 👇 1) AssemblyAI is processing terabytes of audio data every single day - podcasts, meetings, phone calls, radio, TV, broadcast, etc 2) Over 100K developers use the API, resulting in 30M AI inference calls per day 3) Every 6 months the cost has been going down due to the economy of scale and model optimizations 4) There is a ton of interest in streaming use cases (voice bots, agent assist, close captions) as well as non-streaming 5) Since non-streaming models can work bi-directionally, they will always produce higher quality. The majority of users submit non-streaming tasks. 6) They recently launched their newest model called Universal-1 7) Universal-1 can do both streaming and non-streaming. It was trained on 12.5M hours of voice data. 90-93% accuracy in English and 90-92% in French, Spanish and German 8) Today’s AI benchmarks are the Wild West. Industry must use independent 3rd parties for benchmarks. The benchmarking data must be closed source so that companies cannot play the system. 9) Average WER is not a good metric as it’s not representative of real-world user needs 10) WER doesn’t include quality for detecting rare words, alphanumerics, proper nouns, emails, formatting, or context. But these are super important for Speech AI workflows (e.g. summaries) 11) What users care about is not WER but fluency of output 12) AssemblyAI is doing a lot of human evaluations of models 13) They used Google TPUv5 for training Universal-1 14) They will always work to make STT models better, faster, and cheaper. STT market will grow faster once the models improve. New use cases will unlock. 15) In 18mo-24mo models will be much more accurate 16) Currently, AssemblyAI is highly focused on STT and Speech Understanding. TTS and Translation will come over time but not soon. Dylan, thanks for your time and insights 🙏 Full interview here 👉 https://lnkd.in/dbSVzc6G

77

1 Comment

1/ Uncomfortable $10B question for $DELL 😳 Analyst Toni Sacconaghi asked a simple yet brutal Q: Why didn't operating profit change when $DELL added $1.7B worth of servers for #AI services? #earnings #tech 2/ The implication? $DELL's operating margin on #AI servers is effectively zero! 😲 A single Q led to a $10B drop in market cap. Ouch! #investors #stockmarket 3/ Are chipmakers the real #winners of the #AI hype? 🤔 While tech giants chase AI, semiconductor companies rake in profits from high demand. #analytics #bigdata 4/ $DELL's situation reminds us that not everyone benefits equally from new tech trends. 📉 The road to #AI domination is paved with costly investments. #riskvreward #innovation 5/ A $10B question indeed! 💸 One analyst's query exposed potential cracks in $DELL's #AI strategy. Will they course-correct or get left behind? 🔭 #disruptor #futuretech 6/ In the world of finance, simple Qs can have massive impacts. 💥 $DELL's case shows why public companies must be transparent & prepared. #accountability #shareholdervalue 7/ The AI wave is rising, but who'll be the real surf champions? 🌊 Tech titans or component suppliers? An exciting race to watch! #competition #disruption 8/ Perhaps $DELL's AI conundrum serves as a wake-up call. 💡 Pivoting to new tech requires careful planning & execution. No room for rookie mistakes! #strategy #leadership 9/ While some rejoice at $DELL's $10B stumble, others see an opportunity. 👀 A shake-up could spur innovation & realign priorities. #transformation #pivot 10/ In the end, the $10B Q highlights the complexities of the #AI revolution. 🧩 Not every player will emerge a winner, but the game's just begun! #futureisnow #techbuzz

1

1 Comment

really good guidance on how to present to VCs

11

1 Comment

OpenAI today announced their new GPT-4o model. Here are the key points you should know: Omni: The "o" in GPT-4o stands for Omni. OpenAI is now focusing on a cross-functional user experience, integrating text, audio, and images into the same AI pipeline. Unlike previous versions of ChatGPT, where you had to convert images into prompts or text into voice, the Omni model offers a seamless experience. This results in faster and more intuitive interactions. Improved Performance: GPT-4o is faster, a LOT FASTER, and more capable across many tasks. It outperforms previous versions of GPT-4, as well as popular models like Gemini and Claude, in numerous publicly listed tasks and test comparisons. While further research will validate these claims, the progress across text, audio, and other AI applications is impressive. When you don't have to wait, you can iterate on prompts and answers much more quickly. Desktop Version: OpenAI is releasing a desktop version of GPT-4o that can access the context of your desktop and screen. This feature enables various collaborative scenarios where ChatGPT understands the application in use. This is perhaps what Clippy was always trying to be, only this AI is exceptionally more helpful. Another significant development is that AI computation is now happening both on-device and off-device. This is expected to be further enhanced by upcoming technology from Apple, likely through next-generation assistants. Additionally, streaming camera and audio input allows for fluid conversations, interruptions, and continuations without losing context. Despite the technology being only a day old, my initial experience over the past few hours indicates that while the AI revolution is well underway, the UX revolution driven by these technologies is just beginning.

31

2 Comments

Generative AI is outpacing the internet and PCs in adoption speed. Nearly 40% of U.S. adults have embraced this technology in just two years, compared to 20% for the internet and PCs in their early days. This rapid uptake signals a seismic shift in how we work and interact with technology. Are you ready for the AI revolution? #GenerativeAI #TechAdoption #FutureOfWork #AIRevolution #DigitalTransformation

8

1 Comment

OpenAI drops hints for their new model. Marc Andreessen sends $50k to an AI twitter account. A "Miss AI" competition, and Elon's neu thing. Here's what happened this week in AI👇 Reuters released that OpenAI is working on a new model called "Strawberry" It builds on the infamous Q* model which allegedly got Sam Altman ousted. OpenAI is building Strawberry to be a reasoning agent that can perform "Deep Research" https://lnkd.in/g_ju45XM 3 mysterious models appeared in the LMSYS Chatbot Arena. "upcoming-gpt-mini" "column-u" "column-r" Same thing happened before the release of GPT-4o. "im-a-good-gpt2-chatbot" "im-also-a-good-gpt2-chatbot" were found on LMSYS. OpenAI is cooking. https://lmsys.org/ Marc Andreessen from a16z sent $50k to an AI Twitter/X account. He negotiated with the AI-agent in a thread. Crazy. But I'm a skeptic. https://lnkd.in/gJtEW6cN OpenAI drama? Of course. Whistleblowers filed a complaint with the SEC that the NDAs at OpenAI are way too tight. https://lnkd.in/gH-xbaq8 Microsoft and OpenAI decide to not join the OpenAI board because of too much scrutiny from regulators. https://lnkd.in/gfs6PYk3 Elon Musk addressed a tool called “Optimus” during a Neuralink livestream. Claimed that Optimus will enable people who lost speech to talk + connect and use a robotic arm. https://lnkd.in/guZQ5Dpb There was a Miss AI competition with 1500 "contestants" The AIs were judged on looks, social media presence, and utility. WHAT is going on. https://lnkd.in/gHmiUq8i OpenAI internally defined a new five-tier system to track its progress toward AGI. Amazon released a shopping assistant called "Rufus" Rufus will allow you to ask product questions like: - "Is this coffee machine easy to clean and maintain" - "What are customers saying about this products" etc. I like this. https://lnkd.in/gRgbBj86 Meta's largest Llama 3 model will release on July 23. 405 billion parameters. https://lnkd.in/gPgvi8Cx At GrowthAssistant, we're constantly thinking about the future. We place 10/10 trained talent in areas like affiliate marketing, data analytics, SEO, paid social, and more. DoorDash, Hubspot, True Classic, and 150 others use us. 

10

2 Comments

Thanks for sharing these insights. What I am seeing is that orgs test an idea using a proprietary #LLM like #OpenAI, then put it into production with an open source LLM using a development platform like #Llamaindex. That way, private data remains private.

1

Leonard Rodman, M.Sc. CSM® CSPO® PMP®, The Top Chicago AI Influencer, reviews how our www.iAsk.Ai (iAsk.Ai - LinkedIn page) combines the best of *ChatGPT and Google-like results in one* platform, along with unbiased answers based only on authoritative and factual information, all for FREE! https://lnkd.in/gunpEHpY Sam Yagan Smriti Jayaraman Kekre Eddie Lou Sean Chou Samara Mejia Hernandez Peri Ehlman Sonia Sahney Nagar Jessie Casey Jeff Jordan Jeff Cantalupo Eugenio Culurciello Lucas Pasch Brad F. Fritz Colcol Shashidhar Bellamkonda Adrian M. Keaton Nankivil Ezra Galston Octava Singapore Mohini S. Murtaza Bandukwala Diana Tiart Danny Pantuso Josh Graves Zach Kaplan Adam Fridman Kshitij Mishra Naveed Asem Our BETA models are open to try now, and listed at the very BOTTOM of www.iAsk.Ai homepage including our own LLM: https://iask.ai/?mode=beta - Reasoning LLM (scary good accuracy) especially at solving hard logic/AI type of questions or hard HOMEWORK QUESTIONS that neither GPT 4.0 nor Gemini nor Perplexity can. www.ExactGPT.ai - Our own mini-LLM. The most Accurate Q&A AI in the world (when it comes to academic LLM benchmarks for Q&A). It's also updated with information to the last second. Best to make sure you get a FACT based reply. https://iask.ai/?mode=fast - preview of Instant AI results! No waiting! 😊 #AiSearch #iAsk #GenAi #search #Ai #AiStartup #searchengine #AiTool #generativeAI #LLM

25

3 Comments

While #AI agents dominated the conversation in 2024, Konstantine Buhler of Sequoia Capital says "swarms" or networks of AI agents will be the theme come 2025. He joins Caroline Hyde to discuss on "Bloomberg Technology." https://lnkd.in/dPA-MrKq Summary below 👇 🌐 AI as a Revolutionary Trend The speaker compares AI's impact to that of the industrial revolution, emphasizing its status as a defining macro trend. They stress the importance of long-term investments in AI, suggesting that its potential extends far beyond the 2024 and 2025 hype cycles. 🧠 Historical Progression of AI Early AI research focused on basic natural language processing (NLP) and computer vision. From 2020 to 2023, large language models emerged, revolutionizing question answering and other text-based tasks. 2024 was dubbed "The Year of Agents," where AI systems advanced from providing instructions to autonomously completing tasks. 🤖 Applications of AI Agents Cybersecurity: Companies like Expo developed AI agents capable of outperforming top human penetration testers. These agents autonomously identified and reported vulnerabilities in websites. Sales Optimization: Rox, another AI company, introduced agents that augment human capabilities in sales. For example, they assist with research and preparation for client meetings without replacing human relationship-building efforts. 🤝 Human and AI Synergy Rox demonstrated the importance of integrating human expertise with AI agents. Their agent, initially designed to send automated emails, was more effective when allowing humans to approve and personalize communications, leading to a 333x performance boost. 🚀 Future Predictions for AI Agents 2025 and beyond may focus on "swarms of agents"—networks of AI working collaboratively or competitively to achieve complex goals. Industries like healthcare and education are poised to benefit significantly, with AI expected to reduce costs, improve efficiency, and enhance service accessibility. Insights Based on Numbers 🔢 Performance Metrics: Rox's integration of AI with human input improved sales performance 333 times over full automation. 💡 Cross-Industry Impact: Healthcare and education, both plagued by rising costs, are expected to see transformative improvements through AI, with enhanced efficiency and accessibility over decades.

1

Meta just released LLaMa 3.1, the most powerful open source LLM to date, which outperforms GPT-4 and Claude 3.5 on some parameters. This is the biggest moment so far in the open vs. closed debate in AI. One of the big questions for its success is: will developers adopt open source (and specifically Meta) to build new products? Mark Zuckerberg shares his vision for the future in "Open Source AI Is the Path Forward" https://lnkd.in/eqbmx7_x?

27

1 Comment

Supply chain enforcement doesn't seem to be stopping anything. In fact when you place resource constraints around groups of motivated people they will actually out innovate you. #Energy efficiency and #computation performance means a lot and amplifies through a system. This will be interesting to watch. 01.ai trained the #7 model in the world for $3M pre-train cost. And the inference price is $0.14/million tokens! Summary: As Chinese entities do not have access to tens of thousands of advanced AI GPUs from companies like Nvidia, companies from this country must innovate to train their advanced AI models. Kai-Fu Lee, the founder and head of 01.ai, said this week that his company has trained one of its advanced AI models using 2,000 GPUs with just $3 million. “The thing that shocks my friends in the Silicon Valley is not just our performance, but that we trained the model with only $3 million and GPT-4 was trained with $80 to $100 million,” said Kai-Fu Lee (via @tsarnick). “GPT-5 is rumored to be trained with about a billion dollars. […] We believe in scaling law, but when you do excellent detailed engineering, it is not the case. […] As a company in China, first, we have limited access to GPUs due to the U.S. regulations [and a valuation disadvantage compared to American AI companies].” #china #ai #efficiency #llm #usa #gpu #ml Kelly Cohen Zachary Huhn Summer Crenshaw Scott Phillips Lohith Srikanth Pentapalli Carl Fraik Paul Baier Alden Do Rosario Eduardo Ordax Joe Oquist💡 Lori Mazor Monica Turner Christi Brown Dave Hatter Josette Riep Michael Beck, MBA Brian Burke Gregg Burke Cory Warfield https://lnkd.in/d5Kx8N94

22

17 Comments

There’s a Cambrian Explosion of Voice AI Agents. They’re proliferating in these categories: existing business call volume, novel LLM-enabled Voice AI applications, and Voice User Interfaces. 1. Existing Business Call Volume. We’re seeing Voice AI eat the market for business phone calls. Examples are Infer (YC S21)'s lead qualification agents, Domu's customer surveys, 11x's inbound sales agents, the afterhours agents and appointment scheduling agents people are building on Phonely (YC S24), and so on.  2. LLM-enabled Voice AI applications. These are products that weren’t possible before LLMs. My favorite is Boardy, an AI that introduces you to other professionals in its network. Voice AI feels like an artist’s medium. I’ve built a few in this category for fun, like an AI podcast host that interviews you and generates social content based on the interview. 3. Voice User Interfaces. Soon it will become glaringly obvious how ineffective it is to furiously thumb-tap our phones to death. Fluidworks.ai is building a voice interface for product demos and user onboarding. Whispr Flow lets you dictate to any device with a keyboard shortcut. Soon Generative Siri will handily accomplish tasks within third-party phone apps. If you’re building any of these, you need to know what’s happening in your calls. We can help! Check out the link in the comments.

44

6 Comments

𝗪𝗲𝗲𝗸 𝟯𝟰: 𝗕𝗲 𝗮 𝗪𝗬𝗦𝗜𝗪𝗬𝗚 𝗣𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘁𝘆 In the computing world, WYSIWYG is an acronym for 𝗪hat 𝗬ou 𝗦ee 𝗜s 𝗪hat 𝗬ou 𝗚et. This was used to refer to software that allows content to be edited in a form that resembles its appearance when printed or displayed as a finished product. This acronym has become very apt to describe personalities who are 𝗮𝘂𝘁𝗵𝗲𝗻𝘁𝗶𝗰, 𝘁𝗿𝘂𝗲 to themselves, 𝗱𝗲𝘃𝗼𝗶𝗱 𝗼𝗳 𝗮𝗻𝘆 𝗽𝗿𝗲𝘁𝗲𝗻𝘀𝗲, 𝘀𝗶𝗻𝗰𝗲𝗿𝗲 and 𝗵𝗼𝗻𝗲𝘀𝘁. These kinds of personalities are in short supply and much needed in the world today. Till a few years back, many of us had a work persona and home persona – switching from one to the other during the (dreaded!) commute. But with work-from-home and changing expectations of Gen Z, the persona lines are increasingly blurred. Unconsciously many of us were bringing “our whole and authentic self” to work – in other words WYSIWYG. This is beneficial in many ways. When all of us get to know each other’s “persona”, it is the perfect foundation for collaboration and teamwork. You don’t have to second guess the intentions behind actions. 𝗕𝗲𝗶𝗻𝗴 𝗮𝘂𝘁𝗵𝗲𝗻𝘁𝗶𝗰 𝗳𝗿𝗲𝗲𝘀 𝘂𝘀 𝗳𝗿𝗼𝗺 𝗮𝗻𝘆 𝗽𝗿𝗲𝘁𝗲𝗻𝘀𝗲 𝗮𝗻𝗱 𝗮𝗹𝗹𝗼𝘄𝘀 𝘂𝘀 𝘁𝗼 𝗯𝗲 𝗼𝘂𝗿𝘀𝗲𝗹𝘃𝗲𝘀. Trying to fit into a persona that you are not, is self-draining to you and confusing to others. 𝗕𝗲 𝘆𝗼𝘂𝗿𝘀𝗲𝗹𝗳 – 𝘆𝗼𝘂 𝗵𝗮𝘃𝗲 𝗻𝗼 𝗽𝗮𝗿𝗮𝗹𝗹𝗲𝗹 𝗶𝗻 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝘄𝗼𝗿𝗹𝗱 – 𝘆𝗼𝘂 𝗮𝗿𝗲 𝘂𝗻𝗶𝗾𝘂𝗲. And it helps. #authentic #wysiwyg #nopretense #honest #sincere Source: Personal experiences and reading Follow me for weekly updates. #Weeklylearnings #TIL #productive #fulfilling

120

8 Comments