Stepfunction podcast

24:44

Jeff and Seymour kick off 2024 with a discussion of the three phases of generative AI. Phase 1: The launch of ChatGPT in November, 2022. Phase 2: The rise and fall of AI wrapper companies during early- to mid-2023. Phase 3: The current emergence of AI Agents that can automatically chain together multiple steps that drastically change the kinds of products that can be built and the software development work behind them. Links: The Information 's article "OpenAI Shifts AI Battleground to Software That Operates Devices, Automates Tasks" Links from Google's 2018 launch of Google Duplex: Blog post by Google Research Initial 2018 review 2019 retropsective from The Verge Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 18 - OpenAI Over Board 27:28

2 years ago27:28

27:28

We re-examine the swirl of events at OpenAI in November, starting with the Nov 6th Dev Day in San Francisco . Then the tumultuous firing and re-hiring of CEO Sam Altman, President Greg Brockman, and related Board ups and downs. Links: History of OpenAI Board . Board as of Nov 1, 2023: Sam Altman, Greg Brockman, Ilya Sutskever, Helen Toner, Tasha McCauley, Adam D'Angelo Board as of Dec 1, 2023: Adam D'Angelo, Larry Summers, Bret Taylor May 2023 - Reid Hoffman left OpenAI Board and announces investment in InflectionAI . History of impeachments at US Supreme Court . Casey Newton's recent Platformer newsletters summarizing events at OpenAI. Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 17 - Ten years from Her (2013) to Humane 28:55

2 years ago28:55

28:55

Jeff and Seymour discuss interplay among sci-fi movies, speculative fiction, and 'real world AI' through the lens of Spike Jonze's 2013 movie Her . Also some conversation about the upcoming Humane Ai Pin . Links: Biden's Executive Order and related AP article quoting White House Deputy Chief of Staff Bruce Reed on impact of Tom Cruise's recent Mission Impossible movie . Quote : "We become what we behold. We shape our tools and then our tools shape us." Deep learning revolution began getting significant attention in 2012 when AlexNet won that year's ImageNet competition . 2013 interview with Spike Jonze when he describes his early 2000s experience instant messaging with a chatbot similar to ELIZA . Original voice behind Her 's Samantha AI was Samantha Morton ; her role was completely re-recorded by Scartlett Johansson during post-production. Samantha Morton also starred with Tom Cruise in an earlier influential sci-fi film Minority Report . Rabbit might be building an AI-based OS with an investment by Sam Altman . NY Times article about a future hardware device mixing contributions from Altman, Jony Ive 's LoveFrom , ARM , and Masayoshi Son 's Softbank . Humane's website where you can view their product photography and pics from their fashion show. Preview before Humane's November 9th product launch this week. Form factor reminded Jeff a little of the 3rd Gen iPod nano from 2007. Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 16 - Midjourney vs Google SGE vs OpenAI DALL-E 3 23:45

2 years ago23:45

23:45

Seymour and Jeff discuss the recently announced updates from OpenAI, especially regarding image generation in GPT-4 and DALL-E 3. Our ranking of image generation AI's from best to worst: (1) Midjourney, (2) Google Search Generative Experience (SGE), and finally (3) DALL-E. Jeff closes by talking about the recent LLM workshop he conducted for junior high and middle school students. Links: OpenAI announces new voice chat and image features for ChatGPT. DALL-E 3 update . Google Converse aka Google SGE is still better than DALL-E. Midjourney is still the best. Regarding earlier deep learning methods of translating sketches into finished drawings, Jeff was thinking of NVIDIA's GauGAN, based on SPatially-Adaptive DEnormalization (SPADE). 2019 blog post by NVIDIA. Associated paper at arXiv and code at GitHub . From Jeff's workshop: Definitions for the G,P, and T in "ChatGPT" Generative (as in generative AI--see this entire podcast 😉). Pre-trained . Transformer . Meta/FB's Llama2 (7 Billion parameters). Fine-Tuning–one of part of many methods to optimize a base model. See charts in this NVIDIA article . Low-Rank Adaptation: Conceptual article about LoRA at HuggingFace. Original LoRA 2021 paper . May 2023 QLoRA paper . August 2023 LoRA-FA paper . Short Wikipedia description of LoRA . 2019 programmer joke about using Google and StackOverflow . Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 15 - The "Who Is Jeff Hwang?" Test 28:45

2 years ago28:45

28:45

We discuss the September 2023 AI Conf in San Francisco, Anthropic AI and Claude, and how to test LLMs with "Who Is Jeff Hwang?". Links: 2023 Google Pixel Launch Event news from The Verge. 2023 Facebook Connect on Metaverse, Quest VR hardware, and AI. Schedule of speakers and talks at AI Conf in SF, September 2023. Amazon invests $1.25B in Anthropic with option to invest up to $4B. Founders of Anthropic are alums of OpenAI. Better answers for "Who is Jeff Hwang" from www.phind.com , www.pi.ai , Bard from Google, and Microsoft Bing Chat . IBM Video 1 : Simple explanation of fine-tuning and prompting LLMs (start at 2:45) IBM Video 2 : Explanation of Retrieal Augmented Generation (RAG) Hugging Face explainer on PEFT (Parameter Efficient Fine-Tuning) Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 14 - End of Microsoft PC, beginning of Microsoft CP (Co-Pilot) Era 27:10

2 years ago27:10

27:10

(Episode originally recorded September 21, 2023) Jeff and Seymour discuss the Microsoft September 2023 Surface hardware event which was mostly about AI, Llama-2 and open-source LLMs, and Elad Gil's 4 Waves of AI adoption. Links: Round-up of news from Microsoft September 21st event from The Verge . Info about Panos Panay's departure from MS and some rumors (paywalled Business Insider article) Info about Llama-2 from Meta/Facebook and Microsoft and update from InfoWorld - Sep, 2023. Bill Gurley of Benchmark talking about Llama-2 on CNBC . Leaked memo from Google (May, 2023) on no strategic moats and Stepfunction Episode 12 where we discussed it. Competing machine learning software libraries / frameworks: TensorFlow from Google PyTorch from Meta/FB Comparative popularity of TensorFlow vs. PyTorch as of 2023 per Knowledge Hut analysis (scroll to "Difference Between TensorFlow and PyTorch" section). October announcement that Meta is rolling out generative AI tools for advertisers Elad Gil's article about 4 waves of AI adoption and some companies in Wave 2 that we mention: Quora Notion Snap Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 13 - Thoughts on MosaicML, Neeva, and InflectionAI. Plus Apple Vision Pro. 27:20

2 years ago27:20

27:20

(Episode originally recorded June 30, 2023) We're back! Summertime is slow; Stepfunction episodes will appear more regularly again in the fall. Topics: Jeff and Seymour discuss the launch of the Apple Vision Pro in early June. Databricks bought MosaicML in June and comparison with earlier Snowflake acquisition of Neeva in late May. Profile of the Neeva search engine in more optimistic times. Major funding round for InflectionAI which is behind the Pi chatbot . Article has quotes from Inflection AI CEO Mustafa Suleyman, Microsoft CTO Kevin Scott, and Nvidia CEO Jensen Huang. Seymour reports back from a brief autonomous vehicle ride experience with Waymo in Arizona . Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 12 - Leaked Google Strategy Memo, Open Source, AI Value Chains 33:59

2 years ago33:59

33:59

Jeff and Seymour briefly discuss Google I/O which happened last week including the release of PaLM 2 . Most of the episode is spent on the strategic and ecosystem implications of an internal Google strategy memo and the impact of open source. They also discuss AI Supply Chains in the context of a series of MIT blog posts from April. Finally, they draw analogies from earlier eras of tech history including the rise of the PC industry and IBM/Oracle's relationship with Linux in the late 20th century. Links: Jordan Burgess pushes back on Google strategy memo Andrej Karpathy's perspective Andrej's overall view of current LLM landscape which is similar to Jeff and Seymour's four part framework of understanding gen AI: (1) pre-trained foundation model; (2) fine-tuning; (3) whole product / end user UX; and (4) distribution Examples of fine-tuning methods: (1) RHLF ; (2) LoRA ; (3) LLaMA Adaptor Explanation of quantization to make foundation models lighter weight Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 11 - Google vs. Apple vs. Meta AI: Who has the right approach? 26:32

2 years ago26:32

26:32

Jeff and Seymour look briefly at earnings and layoffs from 2023 Q1 for the Big 5 Tech giants which leads to further pessimism about Google BARD. This sparks a larger conversation about how best to organize internal research groups. Which extreme works best–Google or Apple? Links: The Information published a detailed article on April 27th about Apple's struggles with Siri (paywall) Brian Khoon Lee's essay on the merger of Google Brain and DeepMind with a broader perspective of academic incentives from his perspective as a recent Google Brain employee Xerox Parc and Bell Labs were the most famous corporate research labs in the second half of the 20th century Meta AI, formerly known as Facebook AI Research (FAIR) , might strike the best balance between pure AI research and commercial application among the big tech companies Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 10 - BARD Just Blorps It Out 26:43

2 years ago26:43

26:43

Jeff and Seymour mention a framework of 3 personas: (1) end-users; (2) technical builders like EPD teams, i.e., Engineering-Product-Design; and (3) business leaders. Topics discussed: End-user experience: what it's like to use the free and paid versions of ChatGPT, compared with other LLM user experiences like Google's Bard, Microsoft's Bing Chat, and niche LLM's like Phind.com NY Times story from 4/16 about Google and Samsung relationship Background 2021 article on Traffic Acquisition Cost (TAC) payments from Google to Apple CBS / 60 Minutes interview with Sundar Pichair from 4/16 Evolving view on competitive landscape, future of search, assistants, and assistive computing. Google might be in bigger trouble than we thought in February Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org .…

1
Episode 9 - Have You Ever Questioned the Nature of Your Reality? 24:19

2 years ago24:19

24:19

Jeff and Seymour mention audio generation technology which leads to a larger conversation about how we can tell if audio, text, images, and other media is a Deepfake . They end with questions about media gatekeepers, social polarization, and how do we know what's real? Links: Vim is Jeff's favorite text editor and 10 reasons you should learn it . Andrej Karpthy's YouTube video on building a transformer from scratch . Simon Willison on AI empowering him to build more software side projects . Open Letter for 6-month moratorium on generative AI development. Italian regulators use GDPR rules to ban use of ChatGPT in Italy . Daring Fireball "It's Game Over For Deep Fakes" which links toJohn H. Meyer's impressive Steve Jobs chatbot video . Google/DeepMind Wavenet and related 2016 paper . Deepfake of Pope Francis in trendy white puffer coat. Deepfake of Donald Trump on his way to jail. At time of recording (3/30/2023), Trump has been indicted but not yet been arrested or arraigned, so these images are faked. Good article that covers earlier fake/misleading photographs. Includes: (1) Alexander Gardner's famous 1863 Gettysburg photo of a dead Confederate soldier and (2) Robert Capa's 1936 photo of a Spanish Republican soldier mid-fall in apparent battle from the Spanish Civil war. Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 8 - Moore's Law, Exponential Improvements, AI Chips, Emergent Behavior 29:13

2 years ago29:13

29:13

The news that Gordon Moore passed away broke last Friday just as we were recording the previous episode. Today, Jeff and Seymour discuss Moore's legacy, the past and future of Moore's Law, and how advancements in microchips have been the foundation of deep learning for more than a decade. Links: NY Times obituary for Gordon Moore and a remembrance by Walden Kirsch at Intel. Apple's latest iPhone has an A16 Bionic system-on-a-chip with 16 billion transistors, compared to 1993's then ground-breaking Intel Pentium which had 3 million transistors. Articles on Moore's Law and Dennard Scaling Excellent piece by The Computer History Museum in Mountain View, CA. 2018 piece by Micron on Moore's Law and Dennard Scaling (aka MOSFET scaling). 2007 piece in IEEE on 30 years of Dennard/MOSFET scaling (PDF). Types of chips AI Accelerators and Deep Learning Processors (DLP). CPUs , FPGAs , and ASICs . Google's Tensor Processing Units ( TPUs ). Accleration of AI progress above Moore's Law Seymour refers to the chart in this 2021 IEEE Spectrum article . 2022 article in Discover magazine by the Physics arXiv Blog. Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org .…

1
Episode 7 - Bigger than Prometheus? 23:28

2 years ago23:28

23:28

Jeff and Seymour debate the significance of the generative AI revolution relative to the invention of the WIMP interface , the web , and even the Industrial and Agricultural Revolutions . Prometheus refers to the humans learning how to control fire ; is gen AI bigger than that?! Articles referenced: Bill Gates "The Age of AI Has Begun" Thomas Friedman "Our New Promethean Moment" (NY Times) Rodney Brooks on GPT and related LLM technology Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org…

1
Episode 6 - Where Will Generative AI Startups Win and Lose? 26:39

2 years ago26:39

26:39

Jeff and Seymour start with another major week of news, including GPT-4 , Microsoft and Google adding generative AI to their office suites, and LLaMA breaking out into the wild and onto Raspbery Pi and smartphones. Simon Willison's 3-day history of LLaMA developments this past weekend. Potentially bigger news than the dramatic collapse of Silicon Valley Bank. Two relevant articles from Elad Gil: Startups vs Incumbents and Why weren't there any big startup winners from the last 10 years of deep learning? Coca-Cola commercial set in an art museum using Stable Diffusion PCs disrupting mainframes in the 1980s in the context of Clay Christiensen's Innovator's Dilemma and theory of disruptive vs sustaining innovation Runway ML applying Stable Diffusion to generative AI for video Questions and feedback? Let us know .…

1
Episode 5 - Lightbulbs or Lava Lamps? DALL-E in The New Yorker 21:35

2 years ago21:35

21:35

Jeff and Seymour dig further into image generators like Midjourney , DALL-E , and Stable Diffusion . What is the role of the visual artist as generative AI tools continue to evolve? How might it replicate the relationship between a great writer and a trusted magazine editor? Additional links: Adam Gopnik's piece on DALL-E 2 in a recent issue of The New Yorker The Girl with The Pearl Earring by Johannes Vermeer and related movie and novel The Last Supper by Leonardo da Vinci Prediction Machines by Ajay Agrawal, Joshua Gans, Avi Goldfarb at the University of Toronto. Quote about introduction of cheap electric light from page 19 Questions and comments? Talk to us .…

Stepfunction Podcast

1
Episode 4 - Typeface and Cezanne's Card Players 23:15

2 years ago23:15

23:15

Jeff and Seymour discuss the launch of a new marketing company using generative AI called Typeface . This leads to a conversation about copyright consent, compensation, and credit. We discuss how the history of photography, Napster , and Paul Cézanne's The Card Players might help us understand the future of image generators like DALL-E , Midjourney , and Stable Diffusion . Questions and comments? Talk to us .…

Stepfunction Podcast

1
Episode 3 - Elevators Up, Stairways Down 20:30

2 years ago20:30

20:30

Jeff and Seymour use stories and analogies to explain the two main approaches to AI: Bottom-Up and Top-Down. The recent wave of AI success is mostly based on the bottom up path which includes machine learning, neural networks, and deep learning. Related links: The case of the construction worker with a nail in his boot Murray Shanahan's excellent December 2022 paper Talking About Large Language Models Meta / Facebook AI research's September 2020 blog post on Retrieval Augmented Generation and Segmentation DeepMind / Google AI's December 2021 paper on using a Retrieval-Enhanced Transformer (aka Retro ) and a database of 2-trillion tokens for improved LLM capabilities. Questions and comments? Talk to us .…

Stepfunction Podcast

1
Episode 2 - Who Opened The Floodgates? 22:17

2 years ago22:17

22:17

Jeff and Seymour discuss the unexpected impact of ChatGPT and how Bing Chat may not be ready for prime time. Did OpenAI unintentionally open Pandora's Box because they were worried someone else would beat them to it? Plus some reassurances that Sydney is definitely not sentient or emotional. Kevin Roose's article in The New York Times and transcript of his long chat session with Bing / Sydney. Podcast episode detailing the behind the scenes at OpenAI in the weeks leading up to the launch of ChatGPT in November, 2022 as discussed by Roose and Casey Newton on their show Hard Fork . Questions and comments? Talk to us .…

Stepfunction Podcast

1
Episode 1 - What is ChatGPT? 25:38

2 years ago25:38