2024 March

Books

Humans do inspiration; machines do validation.

Math is good at optimizing a known system; humans are good at finding a new one. Put another way, change favors local maxima; innovation favors global disruption.

― “Lean Analytics, Use Data to Build a Better Startup Faster”

Sometimes people who are doing hard data work may forget to step back and look at the big picture. This is a common mistake because we can definitely go from scientific data analysis to actionable insight for making better business decision. But we need to have some additional thoughts about whether the decision is a global optima or it’s just local due to the limited sample, restricted project goal, or restricted team scope.

Articles

When I decided to end my voluntary immersion in the driver community, I could not shake the feeling that the depersonalization of app workers is a feature, not a bug, of an economic model born of and emboldened by transformations that are underway across the global economy. This includes increasingly prevalent work arrangements characterized by weak employer-worker relations (independent contracting), strong reliance on technology (algorithmic management, platform-mediated communication), and social isolation (no coworkers and limited customer interactions).

As forces continue to erode traditional forms of identity support, meaningful selfdefinition at work will increasingly rely on how we collectively use and misuse innovative technologies and business models.
For example, how can companies deploy algorithmic management in a way that doesn’t threaten and depersonalize workers? How can focusing on the narratives that underlie and animate identities help workers reimagine what they really want and deserve out of a career coming out of the pandemic and the Great Resignation? Will increasingly immersive and realistic digital environments like the metaverse function as identity playgrounds for workers in the future? How will Web3 broadly, and the emergence of novel forms of organizing specifically (e.g., decentralized autonomous organizations or DAOs), affect the careers, connections, and causes that are so important to workers? What role can social media platforms, online discussion forums, and other types of virtual water coolers play in helping independent workers craft and sustain a desirable work identity? In short, how can we retain the human element in the face of increasingly shrewd resource management tactics?”

― “Dehumanization is a Feature of Gig Work, Not a Bug”, Harvard Business Review, The Year in Tech 2024 [Article]

This reminds me that last year when I was on vacation in LA, I’ve talked to a driver worked for both Lyft and Uber in LA. He complained Lyft’s route recommendation algorithm is shitty, not helpful at all, a waste of time, while Uber is better in comparison. At that time I realized how important it is to strengthen employer worker relations and gather feedback from workers or clients on the product improvement.
This is a great article where the author raises his concerns of dehumanization of workers in the future. While technology is advancing and economy is transforming, we don’t expect people to forget who they are in their daily basis work.

Bringing a new technology to market presents a chicken-or-egg problem: The product needs a supply of complementary offerings, but the suppliers and complementors don’t exist yet, and no entrepreneur wants to throw their lot in with a technology that isn’t on the market yet.

There are two ways of “solving” this problem. First, you can time the market, and wait until the ecosystem matures— though you risk waiting a long time. Second, you can drive the market, or supply all the necessary inputs and complements yourself.

― “Does Elon Musk Have a Strategy?”, Harvard Business Review, The Year in Tech 2024 [Article]

Here are two examples. To drive the market, Musk supplies both electric vehicles and charging stations. Zuckerberg proposed the concept of metaverse and changed his company’s name.

This is where Musk’s Wall Street critics might say he’s weakest. Many of his businesses don’t articulate a clear logic, which is demonstrated by the unpredictable way these businesses ultimately reach solutions or products.

Musk has spelled out some of his prior logic in a set of “Master Plans,” but most of the logical basis for exactly how he will succeed remains ambiguous. But this isn’t necessarily Musk’s fault or due to any negligence per se: When pursuing new technologies, particularly ones that open up a new market, there is no one who can anticipate the full set of possibilities of what that technology will be able to do (and what it will not be able to do).

― “Does Elon Musk Have a Strategy?”, Harvard Business Review, The Year in Tech 2024 [Article]

Elon is interesting, but I have to say that we human need this type of person to leap to the future.

What could he possibly want with Twitter? The thing is, over the last decade, the technological landscape has changed, and how and when to moderate speech has become a critical problem-and an existential problem for social media companies. In other words, moderating speech has looked more and more like the kind of big, complex strategic problem that captures Musk’s interest.

― “Does Elon Musk Have a Strategy?”, Harvard Business Review, The Year in Tech 2024 [Article]

This is a great article which profiles Elon Musk specifically in his strategies and vision. It answered my question confusing me for two years: what was Musk thinking on buying Twitter? The answer is: Musk’s vision is not in pursuit of a specific type of solution but is in pursuit of a specific type of problem. If we go back to 2016, Igor as a CEO of Disney decided not to buy Twitter because he looked at Twitter as the solution: a global distribution platform, while concerned the quality of speech on it is a problem. Musk was looking for challenges and complexities while Igor was preventing them and looking for solutions.

YouTube

“One of the things that I think OpenAI is doing that is the most important of everything that we are doing is putting powerful technology in the hands of people for free as a public good. We don’t run ads on our free version. We don’t monetize it in other ways. I think that kind of ‘open’ is very important, and is a huge deal for how we fulfill the mission. “― Sam

“For active learning, the thing is it truly needs a problem. It needs a problem that requires it. It is very hard to do research about the capability of active learning if you don’t have a task. You will come up with an artificial task, get good results, but not really convince anyone. “, “Active learning will actually arrive with the problem that requires it to pop up.”― Ilya

“To build an AGI, I think it’s going to be Deep Learning plus some ideas, and self-play would be one of these ideas. Self-play has such properties that can surprise us in truly novel ways. Almost all self-play systems produce surprising behaviors that we didn’t expect. They are creating solutions to problems.”, “Not just random surprise but to find the surprising solution to a problem.”― Ilya

“Transferring from simulation to the real world is definitely possible and it’s been exhibited many times by many different groups. It’s been especially successful in vision. Also OpenAI in the summer has demonstrated robot hand which was trained entirely in simulation. “, “The policy that it learned in simulation was trained to be very adaptive. So adaptive that when you transfer if could very quickly adapt to the physical world.”― Ilya

“The real world that I would imagine is one where humanity are like the board members of a company where the AGI is the CEO. The picture I would imagine is you have some kind of different entities, countries or cities, and the people who live there vote for what the AGI that represents them should do. You could have multiple AGIs, you would have an AGI for a city, for a country, and it would be trying to in effects take the democratic process to the next level.” “(And the board can always fire the CEO), press the reset button, re-randomize the parameters.”― Ilya

“It’s definitely possible to build AI system which will want to be controlled by their humans.”, “It will be possible to program an AGI to design it in such a way that it will have a similar deep drive that it will be delighted to fulfill, and the drive will be to help humans flourish.”― Ilya

“I don’t know if most people are good. I think that when it really counts, people can be better than we think.”― Ilya

Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast [Sam]

Ilya Sutskever: Deep Learning | Lex Fridman Podcast [Ilya]

I watched Lex’s interview with Sam Altman uploaded on March 18, 2024, and an older interview with Ilya Sutskever happened 3 years ago. Elon’s lawsuit against OpenAI frustrated Sam but Sam is optimistic about the future and everything he is going to release in the next few months. Sam answered the questions “what does open mean in OpenAI” that ‘open’ mainly means putting powerful tech in the hands of people for free as a public good, but not necessarily mean open-source. He said there can be open-source models or closed-source models. About the transition between non-profit to capped for-profit, Sam said OpenAI is not setting a precedent for startup to mimic it but he suggested most startups should go for for-profit directly if they pursue profitability in the beginning.

Ilya’s interview is more interesting to me because he talked a lot about vision, tech, philosophy in Deep Learning. It’s impressive that he had such thoughts 3 years ago.

News

Speech is one kind of liability for companies using generative AI. The design of these systems can create other kinds of harms—by, say, introducing bias in hiring, giving bad advice, or simply making up information that might lead to financial damages for a person who trusts these systems.

Because AIs can be used in so many ways, in so many industries, it may take time to understand what their harms are in a variety of contexts, and how best to regulate them, says Schultz.

― The AI Industry Is Steaming Toward A Legal Iceberg [Link]

The Section 230 of Communications Decency Act of 1996 has protected internet platforms from being held liable for the things we say on them, but it doesn’t cover speech that a company’s AI generates. It’s likely that in the future companies use AI will be liable for whatever it does. It could be a driver of pushing companies to take effort to avoid problematic AI output, and reduce “hallucinations” (when GenAI makes stuff up).

A very obvious downside is people could use AI to do harmful things, but I think people are good to work together and prevent that from different angles such as legal aspect or open source software. I worry more about things that are potential or cannot be seen at least in these years when AI is still immature - which is being too early to rely on AI and deviating from truth. For example, it is too soon for people to lose faith in jobs like teachers, historians, journalists, writers, etc, but I’m concerning people are already losing faith in those jobs because of the development of AI and some violations of copyrighted work. As we have seen that AI could have wrong understanding of facts, have biased opinions, and make things up that don’t exist, we lives could fight for the truth but the dead cannot talk.

China’s latest EV is a ‘connected’ car from smart phone and electronics maker Xiaomi [Link]

Xiaomi started EV manufacturing since 2021 and launched its first EV “SU7” on March 28th 2024. It has the following reasons of success: 1) efficient technology manufacturing in a large scale. Though Xiaomi has no experience in auto field, it is a supply chain master, and has perfect partnership with various suppliers. 2) affordable price. SU7’s start price is 215900 yuan while Tesla’a model 3 is 245900 yuan. 3) customer experience oriented innovation. SU7 model can link to over 1000 Xiaomi devices as well as Apple’s devices. In addition, Xiaomi aims to connect its cars with its phones and home appliances in a “Human x Car x Home” ecosystem.

“The world is just now realizing how important high-speed inference is to generative Al,” said Madra. “At Groa, we’re giving developers the speed, low latency, and efficiency they need to deliver on the generative Al promise. I have been a big fan of Groq since I first met Jonathan in 2016 and I am thrilled to join him and the Groq team in their quest to bring the fastest inference engine to the world.”

“Separating GroqCloud and Groq Systems into two business units will enable Groq to continue to innovate at a rapid clip, accelerate inference, and lead the Al chip race, while the legacy providers and other big names in Al are still trying to build a chip that can compete with our LPU,” added Ross.

― Groq® Acquires Definitive Intelligence to Launch GroqCloud [Link]

Al chip startup Groq acquired Definitive Intelligence to launch GroqCloud business unit led by Definitive Intelligence’s CEO Sunny Madra. Groq is also forming a Groq Systems business unit by infusing engineering resources from Definitive Intelligence, which aims to greatly expanding its customer and developer ecosystem.

Groq’s founder Janathan Ross is the inventor of the Google Tensor Processing Unit (TPU), Google’s custom Al accelerator chip used to run models. Groq is creating a Language Processing Unit (LPU) inference engine, which is claimed to be able to run LLM at 10x speed. Now GroqCloud provides customers the Groq LPU inference engine via the self-serve playground.

The House voted to advance a bill that could get TikTok banned in the U.S. on Wednesday. In a 352-65 vote, representatives passed the bipartisan bill that would force ByteDance, the parent company of TikTok, to either sell the video-sharing platform or prohibit it from becoming available in the U.S.

― What to Know About the Bill That Could Get TikTok Banned in the U.S. [Link]

TikTok is considered as critical threats to US national security because it is owned by ByteDance and required to collaborate with the Chinese Communist Party (CCP). If the bill is passed then ByteDance has to either sell the platform within 180 days or face a ban. TikTok informed users that Congress is planning a total ban of TikTok and encouraged users to speak out against the ban. Shou Zi Chew said the ban would put more than 300000 American jobs at risk.

San Francisco-based Anthropic introduced three new AI models — Claude 3 Opus, Sonnet and Haiku. The literary names hint at the capabilities of each model, with Opus being the most powerful and Haiku the lightest and quickest. Opus and Sonnet are available to developers now, while Haiku will arrive in the coming weeks, the company said on Monday.

― AI Startup Anthropic Launches New Models for Chatbot Claude [Link]

Waymo’s progress in California comes after General Motors-owned Cruise and Apple bowed out of the autonomous vehicle business in California, while Elon Musk’s Tesla has yet to develop an autonomous vehicle that can safely operate without a human driver at the controls.

― Waymo approved by regulator to expand robotaxi service in Los Angeles, San Francisco Peninsula [Link]

Elon Musk requires “FSD” demo for every prospective Tesla buyer in North America [Link]

Full Self Driving era seems to start, but Tesla’s FSD system does not turn cars into autonomous vehicles, so drivers still need to be attentive to the road and ready to steer or brake at any time while using FSD or FSD Beta. Will FSD help with Tesla’s stock?

SpaceX Starship disintegrates after completing most of third test flight [Link]

SpaceX’s Starship rocket successfully completed a repeat of stage separation during initial ascent, open and close its payload door in orbit, the transfer of super-cooled rocket propellant from one tank to another during spaceflight. But it skipped Raptor engine re-ignition test, failed re-entry to the atmosphere, and flying the rocked back to Earth. Overall, completion of many of the objectives represented progress in the development of spacecraft for the business and SpaceX and NASA’s moon program.

Open Release of Grok-1 [Link]

Musk founded xAI in March 2023 aiming to “understand the true nature of the universe”. It released the weights and network architecture of 314B Grok-1 on March 17, 2024. It’s under the Apache 2.0 license meaning it allows for commercial use. The model can be found in Github.

GB200 has a somewhat more modest seven times the performance of an H100, and Nvidia says it offers four times the training speed.

Nvidia is counting on companies to buy large quantities of these GPUs, of course, and is packaging them in larger designs, like the GB200 NVL72, which plugs 36 CPUs and 72 GPUs into a single liquid-cooled rack for a total of 720 petaflops of AI training performance or 1,440 petaflops (aka 1.4 exaflops) of inference. It has nearly two miles of cables inside, with 5,000 individual cables.

And of course, Nvidia is happy to offer companies the rest of the solution, too. Here’s the DGX Superpod for DGX GB200, which combines eight systems in one for a total of 288 CPUs, 576 GPUs, 240TB of memory, and 11.5 exaflops of FP4 computing.

Nvidia says its systems can scale to tens of thousands of the GB200 superchips, connected together with 800Gbps networking with its new Quantum-X800 InfiniBand (for up to 144 connections) or Spectrum-X800 ethernet (for up to 64 connections).

― Nvidia reveals Blackwell B200 GPU, the ‘world’s most powerful chip’ for AI [Link] [keynote]

Two B200 GPUs combined with one Grace CPU is a GB200 Blackwell Superchip. Two GB200 superchip is one Blackwell compute node. 18 Blackwell compute notes contain 36 CPU + 72 GPUs, becoming one larger virtual GPU - GB200 NVL72.

Nvidia also offers packages for companies such as DGX Superpod for DGX GB200 which combines 8 such GB200 NVL72. 8 GB200 NVL72 combined with xx becomes one GB200 NVL72 compute rack. And the AI factory or full data center in the future would consists about 56 GB200 NVL72 compute racks, which is in total around 32000 GPUs.

The Blackwell superchip will be 4 times faster and 25 times energy efficient than H100.

OpenAI is expected to release a ‘materially better’ GPT-5 for its chatbot mid-year, sources say[Link]

On March 14 (local time), during a meeting with the Korean Silicon Valley correspondent group, CEO Altman mentioned, “I am not sure when GPT-5 will be released, but it will make significant progress as a model taking a leap forward in advanced reasoning capabilities. There are many questions about whether there are any limits to GPT, but I can confidently say ‘no’.” He expressed confidence that if sufficient computing resources are invested, building AGI that surpasses human capabilities is entirely feasible.

Other news:

Elon Musk sues OpenAI for abandoning its mission to benefit humanity [Link]

A major AT&T data leak posted to the dark web included passcodes, Social Security numbers [Link]

Apple accused of monopolizing smartphone markets in US antitrust lawsuit [Link]

Amazon Invests $2.75 Billion in AI Startup Anthropic [Link]

Adam Neumann looks to buy back WeWork for more than $500M: sources [Link]

NVIDIA Powers Japan’s ABCI-Q Supercomputer for Quantum Research [Link]

Lilac Joins Databricks to Simplify Unstructured Data Evaluation for Generative AI [Link]

Papers and Reports

Scaling Instructable Agents Across Many Simulated Worlds [Link]

Google DeepMind SIMA Team is working on the Scalable, Instructable, Multiworld Agent (SIMA) project. The goal is to develop an agent that follows instructions to complete tasks in any 3D environments. So far they are making progress on making AI agent understand the environment from computer screen, and use keyboard-and-mouse controls to interact with environment, follow language instructions, and play the video game to maximize the win-rate.

OpenAI has similar work called OpenAI Universe, which aims to train and validate AI agent on performing real world tasks. They started from video game environment as well. Although the goals of these two project sound similar, the minor difference is that OpenAI Universe intended to develop a platform where AI is able to interact with games, websites, and applications, while SIMA aims to develop an AI agent or maybe a robot to interact with the real world.

Announcing HyperGAI: a New Chapter in Multimodal Gen AI [Link]

Introducing HPT: A Groundbreaking Family of Leading Multimodal LLMs [Link]

The startup HyperGAI aims to develop models for multimodal understanding and multimodal generation. They released HPT air and HPT pro. HPT pro outperforms GPT-4V and Gemini Pro on the MMbench and SEED-Image benchmark.

Mora: Enabling Generalist Video Generation via A Multi-Agent Framework [Link]

Sora is the first video generation model, however it is not open-source. Lehigh University and Microsoft Research developed Mora to address the gap of no other video generation models to parallel with Sora in performance. Mora introduces an innovative multi-agent framework. As a result, Mora marks a considerable advancement in video generation from text prompts. The evaluation shows that Mora competes with Sora on most of the tasks, but not as refined as Sora in tasks such as changes in the video content, and video connectivity.

Substack

Streaming Giants Earnings - App Economy Insights [Link]

CrowdStrike has repeated in its investor presentations how it wants to be the leading ‘Security Cloud’ and emulate other category-defining cloud platforms:

• Workday (HR Cloud).
• Salesforce (CRM Cloud).
• ServiceNow (Service Management Cloud).

Public cloud software companies are overwhelmingly unprofitable businesses. However, in FY24, Salesforce (CRM) demonstrated that margins can expand quickly once the focus turns to the bottom line (see visual). And when the entire business is driven by recurring subscription revenue and highly predictable unit economics, you are looking at a finely-tuned cash flow machine.

― CrowdStrike: AI-Powered Security - App Economy Insights [Link]

GRANOLAS: Europe’s Darlings - App Economy Insights [Link]

Oracle became TikTok’s cloud provider in 2020 for US users. With the risk of a TikTok ban in America, we’ll look at the potential revenue impact.

Catz believes this growth (of OCI revenue) is driven by:

• Price Performance.
• Full-stack technology for mission-critical workloads.
• AI capabilities focused on business outcomes.
• Deployment flexibility.
• Multi-cloud offerings.

― Oracle: Cloud & AI Focus - App Economy Insights [Link]

Oracle services for enterprise software and cloud solutions: 1) cloud suite (cloud applications and services), 2) data analytics, 3) autonomous database, 4) enterprise resource planning (ERP) to improve operational efficiencies and integrated solutions to streamline complex business functions.

Key news highlights: 1) Oracle acquired Cerner in June 2022 which is a leading provider of electronic health records (EHR) and other healthcare IT solutions used by hospitals and health systems. Oracle is expanding cloud services including the upcoming launch of Ambulatory Clinic Cloud Application Suite for Cerner customers. 2) The adoption of Oracle Cloud Infrastructure (OCI) are across different segments: cloud natives customers such as Zoom, Uber, ByteDance looking for high price performance and integrated security and privacy, AI/ML customers looking for key differentiation, compute performance, and networking design, generative AI customers looking for control, data security, privacy, and governance. 3) TikTok is probably an essential component of the growth of OCI Gen2 infrastructure cloud services. 4) Oracle signed big Generation 2 Cloud infrastructure contract with Nvidia. 5) Oracle is a critical customer in Sovereign AI. It’s starting to win business per country for sovereign cloud, especially the cloud companies in Japan.

NVIDIA ‘AI Woodstock’ - App Economy Insights [Link]

We spent $700,000 on [our five-second Super Bowl ad] in total and yet earned over 60 million social media impressions.” [link]

― Duolingo: Gamified Learning - App Economy Insights [Link]

Duolingo launched the Duolingo Max subscription tier ($168/year), with Gen AI features enabling a more conversational and listening approach. Duolingo has leveraged AI in two areas: 1) using AI to create content, which allows it to experiment faster, 2) using AI to power spoken conversation with characters.

What is coming: Duolingo launched Math and Music courses into its app in 2023.

Articles

OpenAI and Elon Musk [Link]

Read arguments between OpenAI and Elon. Learned that Elon once believed there is 0% probability for OpenAI to succeed and wanted OpenAI to become for-profit so it can be merged to Tesla and being controlled by Elon himself.