OpenAI Makes a Big Mistake

May 19, 2023

A reminder for new readers. That Was The Week collects the best writing on critical issues in tech, startups, and venture capital. I select the articles because they are of interest. The selections often include things I disagree with. The articles are only snippets. Click on the headline to go to the original. I express my point of view in the editorial and the weekly video.

Content this week from: @sama, @bishrar, @stevesi, @natolambert, @jglasner, @joshconstine



In this week's "That Was The Week," my attention was riveted by the discussions around AI regulation, particularly Sam Altman's testimony at the Senate AI hearing. As a staunch advocate for AI innovation, I've always admired Altman’s contributions. However, his recent advocacy for a regulatory licensing regime for AI models gives me pause.

The specter of Facebook's Libra token debacle looms large in my mind. In its quest for global government approval for Libra, Facebook found itself ensnared in a labyrinth of bureaucracy and regulatory hurdles. The fallout was a significant delay in launch and an eventual rebranding to Diem. The Financial Times in November 2020 reported,

“First launched in June 2019, the scaling down of Libra’s vision comes as it has received a skeptical reception from global regulators, who have warned that it could threaten monetary stability and become a hotbed for money laundering.”

Altman's push for a similar regulatory approach could potentially steer OpenAI into a comparable quagmire, hampering its progress with bureaucratic red tape.

Altman proposed,

"Number one, I would form an agency that issues licensing of models above a certain scale of capabilities, and can take that license away to ensure certain safety compliance standards."

While well-intentioned, this approach is misguided and could potentially throttle innovation and progress in the AI industry. Creating a new regulatory agency would inevitably lead to a protracted approval process, obstructing the development and deployment of AI models.

In contrast to Altman's views, Steven Sinofsky offers a more skeptical perspective on government regulation. In his article, Sinofsky argues against the regulation of AI based on compute or model size, labeling it

"the most a-technical and short-sighted way to regulate."

His skepticism towards government regulation, likely shaped by his experience in the tech industry, is right. Sinofsky's article states,

"Regulation needs to respond to real and not hypothetical concerns... Above all it should not be designed around some obvious point in time technology known to grow/change."

Sinofsky's viewpoint underscores the need for a more laissez-faire approach to AI regulation, considering the rapid pace of technological advancement and the potential implications for humanity of slowing it down. If all major breakthroughs had been slowed or stopped at the point they were flawed but promising, much of human progress would never have happened.

Nathan Lambert's article "Unfortunately, OpenAI and Google have moats" delves into the competitive advantages that companies like OpenAI and Google have in the AI space. This is in response to the recent leak of an internal Google document claiming no such moats exist. Lambert underscores the importance of data and user habits in creating a competitive moat for these companies. He also highlights the challenges that open-source AI models face in replicating the success of these companies. This article serves as a stark reminder of the importance of fostering innovation in the AI industry, even as we grapple with the question of regulation.

While I understand Altman's call for a regulatory licensing regime as a defensive response to criticism of LLM AI models, I firmly believe we must resist such regulation. Altman and Sinofsky's contrasting views and Lambert's insights on the importance of innovation underscore the ongoing debates within the tech industry about the role of government regulation. It's clear to me that we need a balanced approach that nurtures innovation while ensuring the responsible use of AI technologies. This can and should come from within the AI industry and does not necessitate Government involvement.

I believe that Chat GPT and others represent an entirely human invention that is capable of enhancing human life, albeit at an early and still flawed stage. Like all tools, the tool is an enhancement to humanity and the division of labor. It is not a thing in itself outside of its human creators and users any more than a hammer is a thing in itself.

Andrew is travelling this week so I have included a video he and I made about UBI and AI in 2020. It is a detailed conversation about how UBI might work in a post-wage-labor society. I would add a lot to it today now Worldcoin has emerged, but it is still a reasonable overview of how we might live in the post-AI future.

Essays of the Week

Globalism is, in fact, Regionalism?

A wonderful new book argues 'twas ever thus. Now, more than ever, it is where the big opportunity lies.


Christopher’s Newsletter
Globalism is, in fact, Regionalism?
As I have written here before - as software has already eaten the world and AI is devouring it even faster - American technology business needs a new global play book. For decades, America pretty much won by showing up. The Google, Intel, Apple, WhatsApp, Instagram and on and on of any country was, well, Google, Intel, Apple, WhatsApp, Instagram etc…
Read more

As I have written here before - as software has already eaten the world and AI is devouring it even faster - American technology business needs a new global play book.

For decades, America pretty much won by showing up. The Google, Intel, Apple, WhatsApp, Instagram and on and on of any country was, well, Google, Intel, Apple, WhatsApp, Instagram etc.

As I also have written at length, the most interesting trend in technology is not the tech itself but the global access to it. As more than half the world walks around with a supercomputer in their pocket, no surprise that juggernaut enterprises are rising everywhere because talent is, also, everywhere. And customers want not just one-stop-shop American answers but solutions that cotton to their own experiences on the ground.

Uber was a classic example of the old American playbook. They had the model, the talent, the data, the tech, and the money to land in any country and take it all. Grab in Southeast Asia, Didi in China, and Careem in the Middle East, however, had different ideas.

For me, and through software, this was an unleashing of a very new kind of globalism. As has long been happening in the trade of physical goods, every company as a software company can “go global” in a fraction of the time and money. Netflix and Amazon Prime stood up their services in days in over 100 countries and the more data they collect the more relevant they are to the ground. What is true of Netflix and Amazon Prime could be true for any enterprise of scale -- and now from anywhere.

I was stopped in my tracks recently, however, in getting to know Shannon K. O’Neil, Nelson and David Rockefeller Senior Fellow and Deputy Director for Latin American Studies at The Council on Foreign Relations — and one of the great minds on trade and supply chains.

What one really needs to pay attention to, she argues, is that globalism is - now more than ever, but always has been - regionalism.

The data is with her and surprised me. Of 500 leading international companies, two out of three of every dollar of revenue came from within their own region. In fact, in total, over half of the flows of international trade, investment, money, information and people happen within regions. Walmart has more stores abroad than in the US, but over 90% of its revenue still comes from North America. Vodaphone is in over two dozen countries, but operations in Europe remain over three quarters of their revenue and profit.

O’Neil digs deeply into her views in a marvelous, must-read new book - breezily written but jammed packed with data and analysis, The Globalization Myth: Why Regions Matter.

The cause of regionalism, she shows, isn’t so surprising. Language and cultural cues remain substantial both among customers and within one’s own corporate culture; manufacturing coordination remains complex; shifting regulation regimes bring consistent risk and more. She tours the world in three parts -- Europe, Asia and North America and – and shows repeatedly that where there is success there is a story of regionalism.

But isn’t America’s reliance on China for manufacturing transcendent of region? She notes that truth, especially in assembly of sensitive technology, but at the same time: “The real story is the regional one. Much more of the international money pouring into factories, warehouses, labs and R&D centers, powering cities and raising skylines, came from Asia than from anywhere else. By the early twenty-first century, China was importing more goods from its neighbors than from the United States and Europe combined.”

Venture-Fund Returns Show Worst Slump in More Than a Decade

Portfolio markdowns are denting the performance of an industry that has long outperformed other asset classes

By Berber Jin

May 16, 2023 9:00 am ET

Source: PitchBook Data

Venture-capital-fund performance is languishing amid the broader downturn for tech startups, denting returns for university endowments, pensions and other investors that increased their exposure to the sector during the bull market. 

For the first time in more than a decade, returns for venture funds were negative for three consecutive quarters last year, according to research firm PitchBook Data, as investors finally began to mark down startups that had ballooned in value. Initial data for the fourth quarter also show a negative quarterly return. 

The data also show that the yearly internal rate of return hit minus 7% in the third quarter—the latest data available for that measure—the lowest value for those three months since 2009. The internal rate of return is used to measure the profitability of venture funds on an annual basis and is a key performance metric used by the industry. 

The decline marked the fifth consecutive quarter of deteriorating yearly rates of return—the first time this has happened in a decade—and was also the only negative rate of return among seven investment categories tracked by PitchBook, including private equity and real estate. 

Fund investors say they remain optimistic about the long-term potential of the venture industry, citing new areas of technological growth like artificial intelligence. Even with last year’s decline, venture-capital funds outperformed other private investment strategies when measured over periods of three years or longer, the data show. 

But the severity of the recent declines is already pushing some investors to re-evaluate their exposure to the sector. 

These fund investors, known as limited partners, are bracing for further markdowns and say that the funds that invested the bulk of their cash during the peak of the bull market will likely post subpar returns. 

Infographic showing the World's Top 50 Largest Endowment Funds

Ranked: The World’s Top 50 Endowment Funds

What do Harvard, the Church Commissioners for England, the NYC Metropolitan Museum of Art, and an entity on behalf of Saudi Arabia’s King Abdullah all have in common? They all have endowment funds.

An endowment fund is the investment arm of nonprofit institutions like universities, charities, and churches. The purpose of the fund is to invest the organization’s assets to fuel future operations and other important projects.

The world’s largest endowment funds have billions in investable assets, making them sizable players in the finance sector. Here, using data from Sovereign Wealth Fund Institute, we take a closer look at the world’s largest endowment funds by total assets.

Types of Endowment Funds

Overall, there are four main types of endowment structures.

  • Unrestricted Endowment: A fund structure where assets are used at the full discretion of the institution

  • Term Endowment: A fund structure with a fixed term time period before the principal can be spent

  • Quasi Endowment: A donation to an endowment with a specific purpose to deploy that capital

  • Restricted Endowment: A fund structure where the principal value from donations is held forever and only returns generated on the principal can be used

In addition, each endowment fund has different structures in regards to withdrawals, use of funds, and their general investment philosophy.

The Largest Endowment Funds

The largest endowment funds can be compared on a grand economic scale, in terms of assets.

To put it all into perspective, the largest 50 endowment funds represent over a trillion dollars in assets. Or for a more singular example, look at Harvard’s fund, which has an endowment greater than the entire GDP of countries like Serbia, Bolivia, or Slovenia.

The largest endowment fund, Ensign Peak Advisors, is based in Salt Lake City, Utah, and manages the assets for the Mormon Church (officially known as the Church of Jesus Christ of Latter-day Saints). The church itself has over 16 million members worldwide and is the fourth largest church in America.

Video of the Week

AI of the Week

Sam Altman Suggests AI Licensing Regime In Senate AI Hearing

Today's senate hearing was a somewhat nuanced discussion in which everyone present kept saying we need to regulate ai


Pirate Wires
Sam Altman Suggests AI Licensing Regime In Senate AI Hearing
In a Senate Judiciary Committee hearing today, OpenAI’s Sam Altman advocated for the establishment of a capacity-based, regulatory licensing regime for AI models that, ideally, larger companies like Google and OpenAI would be subject to, but smaller, open-sourc…
Read more
Sam Altman testifying at the Oversight of AI: Rules for Artificial Intelligence Senate hearing

In a Senate Judiciary Committee hearing today, OpenAI’s Sam Altman advocated for the establishment of a capacity-based, regulatory licensing regime for AI models that, ideally, larger companies like Google and OpenAI would be subject to, but smaller, open-source ones would not. “Where I think the licensing scheme comes in is as we head toward AGI — I think we need to treat that as seriously as we do other technologies,” Altman said.

When Sen. Graham (R-SC) suggested a government agency “that issues licenses and takes them away,” Altman agreed.

In addition to Graham and Blumenthal, the hearing included questioning from Josh Hawley (R-MO), Mazie Hirono (D-HI), Cory Booker (D-NJ), Dick Durbin (D-IL), and several other senators. Christina Montgomery, VP/ Chief Privacy & Trust Officer for IBM and NYU professor Gary Marcus also provided testimony. All speakers — witnesses and senators alike — agreed that AI should be regulated, though Altman, Montgomery, and Marcus differed on specifics, and senators like Hirono questioned if effective licensing was even possible.

Christina Montgomery testifying at the Oversight of AI: Rules for Artificial Intelligence Senate hearing

Of the witnesses, IBM’s Montgomery seemed to have the most cautious position. “IBM urges congress to adopt a precision regulation approach to AI,” she testified. “This means establishing rules to govern the deployment of AI in specific use cases, not regulating the technology itself. Such an approach would involve four things: First, different rules for different risks. The strongest regulation should be applied to use cases with the greatest risk to people and society. Second, clearly defining risks. There must be clear guidance on certain uses or categories of AI-supported activity that are inherently high risk. Third, be transparent. AI shouldn’t be hidden. Consumers should know when they’re interacting with an AI system. Finally, showing the impact. For higher risk use cases, companies should be required to conduct impact assessments that show how their systems perform against tests for bias, and other ways that they could potentially impact the public.”

“AI should be regulated at the point of risk,” Montgomery said. Later she indicated her position was that some uses of AI should be licensed, but not all.

OpenAI’s Altman suggested more aggressive regulation. “Number one, I would form an agency that issues licensing of models above a certain scale of capabilities, and can take that license away to ensure certain safety compliance standards. Number two, I would create a set of safety standards, specific tests that a model has to pass before it can be deployed into the world. And third, I would require independent audits by experts who can say the model is or isn’t in compliance with safety thresholds.”

206. On Congressional Hearings On Artificial Intelligence (May 2023)

Follow the science…fiction?


Hardcore Software by Steven Sinofsky
206. On Congressional Hearings On Artificial Intelligence (May 2023)
In the AI hearings a panelist suggested the best approach to regulation would be to regulate based on compute or model size. This seems to be the most a-technical and short-sighted way to regulate. 1/8 2/ imagine it is the early 1970s and fear of “databanks” (SQL) sweeps across the industrialized world. In 1980 disk drives held 30MB. Imagine the downstre…
Read more

In the AI hearings a panelist suggested the best approach to regulation would be to regulate based on compute or model size. This seems to be the most a-technical and short-sighted way to regulate. 1/8

2/ imagine it is the early 1970s and fear of “databanks” (SQL) sweeps across the industrialized world. In 1980 disk drives held 30MB. Imagine the downstream effects of capping / regulating drives over 30MB?

3/ How would this have changed the innovation trajectory? Would anyone have kept making large drives? Would the only government have access to large drives? Would we ever have 100,000 of our own photos? Even CD-ROM?

4/ A cynic would say this is just the incumbent trying to define the market / regulatory boundary to their advantage as it exists today. We saw this happen with energy companies. We saw this with healthcare/pharma. With transportation.

5/ this is a shelf of books from the 80s on risks of computers and privacy. Only at the highest altitude were these predictions close. But imagine if we capped network bandwidth, storage, or processing power to “prevent” these risks?

DAVID BURNHAM THE RISE OF THE COMPUTER STATE Miller The Assault on Privacy JOHN WICKLEIN ELECTRONTO NIGI Databanks in Noiand Acodemy diScences a Free Society VIKING Westinand Bake NYT NATIONAL RESEARCH COUNCIL trust in cyberspace NATIONAL ACADEMY PRESS FATAL DEET Ivars Peterson FUNDAMENTALS OF DATA PROCESSING • LYTEL •, FOULSHAM-SAMS The Invasion of Privacy MANGWOKA SATIO NES Privacy, Computers and You The Death of Privacy RHEINGOLD The VIRTUAL COMMUNITY Honestoading on the Elodronic Front The Computer and Invasion of Privacy
Some books I own from the 1980s or so on the risks of databases and privacy.

6/ Everyone, no matter where you are on any issue, should be wary of when industry leaders call for regulation that happens to align [sic] with what they might claim is a competitive advantage in other contexts at this moment.

7/ one question I wish would get asked more is what hypothetical concerns about AI usage are not already covered by existing regulations? AI does not exist nor is it used independent of any system, just as storage isn’t about privacy itself.

8/ Regulation needs to respond to real and not hypothetical concerns (v theoretical concerns based on physics as we see in construction or based on biology as we see in pharma). Above all it should not be designed around some obvious point in time technology known to grow/change.

PS:/ Full link to hearing and transcript here.

Unfortunately, OpenAI and Google have moats

While everyone went crazy over a leaked memo, no one took the time to think through how companies have worked in the internet era.


Unfortunately, OpenAI and Google have moats
The companies that have users interacting with their models consistently have moats through data and habits. The models themselves are not a moat, as I discussed at the end of last year when I tried to predict machine learning moats, but there are things in the modern large language model (LLM) space that open-source will really struggle to replicate. C…
Read more

The companies that have users interacting with their models consistently have moats through data and habits. The models themselves are not a moat, as I discussed at the end of last year when I tried to predict machine learning moats, but there are things in the modern large language model (LLM) space that open-source will really struggle to replicate. Concretely, that difference is access to quality and diverse training prompts for fine-tuning. While I want open-source to win out for personal philosophical and financial factors, this obviously is not a walk in the park for the open-source community. It'll be a siege of a castle with, you guessed it, a moat. We'll see if the moat holds.

To set the stage, in that article I explained what a moat is:

In old-school technology companies, an important defining factor is how they can create a moat protecting investment from their (potential) competitors. Moats often involve technological advantages, users, data, and, most importantly, feedback loops. Feedback loops are what make users continue using the technology and give the companies the resources to make it an advantage more solid. Some modern examples include Office's incumbency for Microsoft, network effects at Facebook or Google, economies of scale at Amazon, branding and proprietary technology at Apple, etc.

At all points along the way, training a chatbot based on large language models is extremely data intensive. If you write out the steps in terms of major technical milestones, it’ll include (simplified):

  1. Train a base multilingual multi-task language model.

  2. Fine-tune the model to answer instructions. People are hacking together solutions for this step (a classic in NLP)

  3. Fine-tune the model further with RLHF to match user values and interests. This is the fine-tuning step I refer to above, extremely new to NLP.

When you hear about recent developments in open source models, it is driven primarily by the release of a better base model (step 1) with a sprinkle of creativity in the initial fine-tuning to make the model answer questions (rather than just predicting the next token ad nauseam. From there, the second (and especially the third) component takes entirely new datasets, engineering infrastructures, and expertise. All three of these are expensive and have been primarily developed behind the closed doors of for-profit companies.

Some numerical rules of thumb that are discussed in this space (they're roughly based on the InstructGPT and Anthropic papers) are that step two takes at least 10k written responses for instruction tuning and step three takes at least 100k comparisons to train a useful reward model and run RLHF (the exact split between reward modeling and RLHF is not clear known). Essentially, once you have the pretrained language model, the next step is to make 6 dataset splits for training for a model with RLHF (see table below from InstructGPT, a now relatively old method). You'll have test+train splits for instruction tuning, reward model training, and RL optimization. The details of best practices for managing each of these distributions are very unknown and it's expected that the last two data subsets in the RL part should be continually updated as model use cases are understood.

A table showing the dataset distributions (length and category) for InstructGPT. I suspect things are much bigger in the number of entries and lengths now.

While these numbers don't sound like a lot on the first pass when these models are trained on trillions of tokens of data from the internet. The challenge is that the distribution and quality of these must match a very narrow distribution and form. These prompts can almost never be scraped. For example, we've examined some open-source datasets on prompts given to ChatGPT (e.g. from ShareGPT), and a generous interpretation of a "useful" prompt leads to only about a 50% acceptance rate from the source data. On top of the inherent quality challenge, some extremely open undocumented datasets like this are barely cracking the 100k training samples region when filtering out obvious not useful prompts.

Now, compare these orders of magnitude to how many queries places like Google and OpenAI are getting. I would posit that both of these organizations may get 100k prompts entered per day. Obviously, the filtering challenges still exist, but that is a huge advantage. Additionally, if OpenAI were to open source the prompt datasets alone, they have more information associated with it that they surely won't release: user behavior data. There will be buckets of prompts that are from "extremely high-quality users" and things that are likely way easier to filter or tune to be useful.

There are already some criticisms of this analysis, such as the Open Assistant dataset. This dataset is very impressive — it's the most performant for instruction tuning that we have found. The data distribution here is almost certainly very different from those used to successfully RLHF at scale. I suspect we will figure out some differences, and they're likely to be addressed, but I don't think the open-source data will be even 75% of the quality of the closed-source counterparts. There are controls in place for the data curation pipelines these companies are paying millions of dollars for — it's not throwing money into the void.

Second, there is simply consumer bias and association moats. The general rule of thumb is that it takes an experience being about 10x better for the average user to switch away from their existing habit. Even if Bing was twice as good as Google, most people would not change. The consumer economy is not driven by tech workers in the hyper newsletter-obsessed news cycle. OpenAI is multiple steps along the path towards ChatGPT being the default word that people use to mean "talk to an AI," much like "Google it" will still be said for decades after Google goes out of business. It amazes me how much people discount this.

What Lightspeed Is Reading, Listening To, And Thinking About AI

Check our our AI Reading List guide for May 2023


Published in Lightspeed Venture Partners

At Lightspeed, AI is not a fad. We’ve been investing in startups with artificial intelligence-based products and services for over seven years, with 30 companies in our portfolio, including Snorkel, Stability AI, and Tome, and we’re tracking dozens more.

Successful platform shifts always rest on a set of underlying foundational primitives and abstractions built to support the new paradigm. The AI revolution is no different. AI’s core building blocks have been stacking up for some time. Now however, the demand side has exploded. And we’re primed to seize the moment with our portfolio companies, present and future.

Lightspeed has placed over $850 million in service of these companies and their mission to build the future. Our investment decisions are based on deep research and analysis, candid communication, and establishing trust and common purpose with the founders we back.

To stay on top of this rapidly evolving industry, Lightspeed’s Partners read deeply and broadly across the field, taking in everything from the technical to the theoretical to the philosophical.

It’s impossible to be comprehensive, but here we’re sharing some of the reads we’ve found most compelling over the past few months. We’ve included some classic texts read in the context of a new excitement, and recent articles explaining the latest developments and dilemmas painting the AI landscape in general and in venture capital. We’ve also included several great aggregators across industry and academia that are seeing around corners in their coverage and analysis.

Foundational Books

An AI Classic Tome
Godel, Escher, Bach
, by Douglas Hofstadter, a researcher in computer science. This book might have been published in 1979, but the questions it asks, like how consciousness emerges from darkness, and how similar thought patterns arose from three geniuses in diverse fields, are newly relevant. If you don’t have time to read a doorstop right now, read these posts on Mind Matters and Medium for a summary (or perhaps to whet your appetite).

The AI Long View
In Artificial Intelligence: A Guide for Thinking Humans, Melanie Mitchell offers a comprehensive and pragmatic overview of the fundamentals of modern artificial intelligence. The book is a tour of the science, a biography of the creators of the technology, a history of invention, and a gaze into the future.

Who is Training Who?
One of the world’s leading academics on AI, Stuart Russell, offers up Human Compatible: Artificial Intelligence and the Problem of Control an accessible primer on the risks that AI poses to humanity, and how we can approach the technology more responsibly.

AI x Startups, Strategy, Business

Do AI Moats Exist in SaaS?
In this ACQ2 podcast episode hosts Ben Gilbert and David Rosenthal talk through b2b AI moats with Jake Saper, General Partner of Emergence Capital. Jake highlights that the likely moat for most b2b AI businesses will be not the LLM itself but the “good old fashioned building blocks of SaaS, like complex workflows, data integrations, advanced permissioning.” This “scaffolding” will create defensible generative AI-enabled businesses.

The AI Litmus Test
VC NFX argues there is a way to determine whether startups are building a “hard, unique, defensible” product. Read on for Morgan and Drew Beller’s hypothesis and a Donald Rumsfeld callout.

Frothy or Brilliant? got 170 offers for funding by running an extremely open fundraise, according to The Information. Can AI also change VC? CEO Dan Siroker also offered a closer look behind the already open curtain on Twitter.

AI McKinsey?
The New Yorker
posits an interesting idea — AI as a management consulting firm, with the capacity to improve business, or do harm, based on inputs. The writer argues the next stage of capitalism will be based on how well “the consultant” is instructed in what its aim should be in helping a business evolve.

The ChatGPT MC
Sam Altman ignited a spark with ChatGPT’s release from his startup OpenAI. This recent profile chronicles his view of what AI can and perhaps will be, and in a Rosebud moment, includes a mention of a short story that is one of his favorite views of artificial intelligence, a sci-fi piece called The Gentle Seduction. (He also argued for AI regulation in a Senate hearing today.)

Does OpenAI have it wrapped?
Spark Capital VC Fraser Kelton writes that
he doesn’t, “see how there’s anything but a small number of groups providing the largest, most capable models. At this point it seems clear that it will be OpenAI and Anthropic and then tbd on whether anyone else will join them.” Provocative! Here’s the leaked Google memo Kelton is responding to.

Pitchbook on Generative AI
Get a free look at what the research firm’s analysts think about the directions this emerging technology can go with their newly released report.

Regulation is Inevitable
In the New York Times Opinion section, FTC Chairwoman Lina Khan lays out a framework for AI regulation: she argues her commission is best equipped to insure equitable use of AI that limits fraud and abuse, and prevents a repeat of the worst excesses of the web 2.0 era.

Better minds, lives, and educators

With ChatGPT, a Runner is Born
This Twitter thread reminds us that prompt engineering is a thing of the present and future. writer Greg Mushen used OpenAI to make himself a running addict, and lost 26 lbs. in the process.

AI for Education
Khan Academy founder Sal Khan argues in his Ted Talk, AI for Education that, “We’re at the cusp of using AI for probably the biggest positive transformation that education has ever seen.” (Here come AI tutors and teaching assistants.) Khan says educators and startups should focus on creating positive tools, not on enabling or detecting cheating.

AI for Mental Health
Can a chatbot replace a therapist? Startup Wysa aims to find out, but as with all health related ventures, patient safety and privacy are big concerns that require definitive answers. NPR looks into it. (Singles might need all the therapy they can get once they get a look at what AI dating might be like.)

Is AI Actually Composability?
Notboring’s Packy McCormick
makes a case that what the new generation of AI tools really do is allow for composing knowledge and complex problem solving. They’re not so much an emergence of intelligence as an alphabet of complex ideas that can be arranged to derive meaning, just as actual alphabets help humans turn the physical world into ideas, concepts, and communication.

On That Note…
Computer scientist and philosopher Jaron Lanier argues in The New Yorker that “There is No AI” in the way it’s being conceived in the popular imagination — as something that has the potential to become an equal — or a competitor — to the human mind. What is being called AI right now are just complex tools — that like all human creations, can be used for good or for harm.

Do Androids Dream of Electric Money?

The AI Investor
Quant firm AQR backtested ChatGPT as an investment tool, and found using it would have outperformed the market from 2004 to 2019, according to a story in Institutional Investor. Unfortunately ChatGPT has only been trained on data up to 2021, and it’s forbidden to dispense investing advice. But you never know!

Fintech meets Generative AI
Bain Capital Ventures
thinks that the fuzzy logic of AI has a role in the exacting world of Fintech, once the kinks are worked out. SaaS, meet GaaS.

Bloomberg Built a LLM
BloombergGPT is the company’s attempt to automate financial tasks, with a bespoke 50 billion parameter Large Language Model. Here’s the abstract and whitepaper.

Academic Research and Resources

NEJM AI in Medicine
Stay up to date on AI x Medical with the leading journal in the space, which is also soon spinning off a new journal focused on AI. More info here.

The Stanford HAI 2023 AI Index Report
“The annual report tracks, collates, distills, and visualizes data relating to artificial intelligence, enabling decision-makers to take meaningful action to advance AI responsibly and ethically with humans in mind.” And trust us, it is chock full of insights. From Human-Centered Artificial Intelligence, which also offers a comprehensive reading list of books about AI. (More from Stanford: Center for Artificial Intelligence in Medicine & Imaging.)

AI Aggregators

Byte Size Updates
Ben’s Bites
is a daily dose of what’s going on across AI. With over 90,000 readers across tech, including from Google, a16z, Sequoia, Amazon, Meta, and more, it’s known for wit and brevity.

A Faire List
Dan Hockenmaier of Lightspeed portfolio company Faire is keeping his own reading list of AI strategy, research, and papers, and it’s worth a bookmark or signup.

Work Smarter With Generative AI
The Lore weekly newsletter is one of our favorite overviews of the weekly key advances, products, and funding rounds in generative AI — with a focus on interactive consumer media and gaming applications.

A Newsletter for Nerds
The Batch
is another great resource established as part of ML researcher Andrew Ng’s, with a focus on research and the economics of LLMs that are core to AI’s functions and capabilities.

And finally, the bizarre
Text to video AI is in its infancy, as this Tweet of an alternatingly gross and perfect pizza commercial made with Runway ML Gen-2 shows. Pepperoni Hug Spot, here we come. Are you ready for best pizza of life?

A Bunch Of AI-Related Companies Are Going Public Via SPAC

May 18, 2023

Joanna Glasner @jglasner

Not a lot of startups are going public lately. But nonetheless, we are seeing a few pursuing market debuts, including a handful of artificial intelligence-related companies taking the risky SPAC route to market.

So far this year, several AI-focused companies have announced tie-ups with SPACs, or special-purpose acquisition companies. Planned mergers span sectors including education, diagnostics and data management.

A few examples:

  • iLearningEngines, a training provider which describes its focus as “AI-powered learning automation,” announced in late April that it plans to list on Nasdaq at an initial valuation of around $1.4 billion through a merger with shell company Arrowroot Acquisition Corp.

  • Spectral MD, which uses artificial intelligence to predict how wounds will heal,  announced last month that it is going public via a merger with a SPAC, Rosecliff Acquisition Corp I. The deal sets an enterprise value of around $170 million for the Dallas- and London-based company.

  • Airship AI Holdings, a Redmond, Washington-based developer of  an AI-driven video, sensor and data management platform, announced in March that it is going public through an acquisition by shell company BYTE Acquisition Corp. at a valuation of around $290 million.

In addition, there are some AI-affiliated blank-check companies that have yet to identify an acquisition target….

News Of the Week

The Breakdown

Section 230 Just Survived a Brush with Death

May 18, 2023 18:54 ET

The U.S. Supreme Court opted not to throw the internet as we know it into utter chaos—for now

By Nabiha Syed

Illustration of the Google logo and Twitter logo standing in front of the U.S. Supreme Court building, looking quite scared.

For those of you worried about the Supreme Court breaking the internet, you can breathe easy. The court left Section 230 of the Communications Decency Act unscathed—for now—in opinions released today on two closely watched cases that many observers worried could shake the foundations of online discourse had they gone differently. 

In a pithy three-page opinion, the court vacated Gonzalez v. Google. The case explored whether Google was liable for acts of terrorism perpetrated by the Islamic State group in Paris in 2015 because the group used YouTube to spread violent messages. A related case, Twitter v. Taamneh, examined whether arguing that online platforms were responsible for the effects of violent content posted by the terrorist group. In a more lengthy opinion authored by Justice Clarence Thomas, the court unanimously found that platforms are not liable under the Antiterrorism Act. Section 230, one of the more important legal provisions of the modern internet, has escaped intact. However, there are a number of interesting wrinkles to consider here, including some hints at where the next challenge to Section 230 may arise.

Report Deeply and Fix Things

Because it turns out moving fast and breaking things broke some super important things.

First, a quick recap: Back in February, the court entertained oral arguments in both cases, taking a close look at liability for choices made by internet platforms. As Kate Klonick, law professor at St. John’s University and fellow at the Berkman Klein Center at Harvard, explained brilliantly here, the narrative arc of Taamneh and Gonzalez should be understood in political context—specifically, partisan calls for the need for social media regulation

Nevertheless, the question at the heart of Gonzalez—whether Section 230 protects platforms when their algorithms target users and recommend someone else’s content—prompted a flurry of concern and an avalanche of amicus briefs discussing why this would break the internet as we know it. (In a Q&A for us, James Grimmelmann, professor at Cornell Law School and Cornell Tech, explained how disruptive this would be for generative AI, too.) Today, the court punted on the case in its opinion, saying it would be resolved by the court’s logic in Taamneh

Taamneh looked at whether Twitter’s failure to remove certain Islamic State content constituted “aiding and abetting” a terrorist attack against a nightclub in Istanbul. The court rejected Taamneh’s claim, explaining that aiding and abetting constitutes “conscious, voluntary, and culpable participation in another’s wrongdoing.” In lay terms, it has to be specific and intentional. Justice Thomas, writing for the court, reasoned by analogy to earlier technologies: “To be sure, it might be that bad actors like ISIS are able to use platforms like defendants’ for illegal—and sometimes terrible—ends. But the same could be said of cell phones, email, or the internet generally.” 

Fascinatingly, this opinion was authored by Thomas, the very same justice who’s been clamoring for increased regulation of social media. Grimmelmann weighs in:

Even more interesting? Justice Ketanji Brown Jackson filed a brief concurrence in Taamneh, noting that “Other cases presenting different allegations and different records may lead to different conclusions,” and in deciding today’s cases, the court “draws on general principles of tort and criminal law to inform its understanding… General principles are not, however, universal.” 

I sent Kate a Signal message to read the tea leaves from Justice Jackson, and what that might mean for the future of Section 230:

So while tech industry lawyers might sleep a bit easier tonight, there’s still more to come. Stay tuned for more internet law intrigue in the coming months as we wait for the Solicitor General’s perspective in the NetChoice cases around social media regulation laws in Florida and Texas. Briefing in those cases is likely to come in the fall, and The Markup will be here to help you make sense of that.

Elon Musk Announces His Pick for the Toughest Job in the World

Linda Yaccarino is leaving NBCUniversal to become the new CEO of Twitter, acquired for $44 billion on October 27 by the billionaire.

LUC OLINGA, MAY 12, 2023 6:10 PM EDT

Elon Musk is once again in the headlines. 

The  Techno King caused a real earthquake on May 11 by announcing to everyone's surprise that he had found a CEO to succeed him at the head of Twitter. 

Three months ago, the billionaire said he would remain boss of the microblogging site until the end of the year.

"I think I need to stabilize the organization and just make sure it's in a financially healthy place and that the product roadmap is clearly laid out," Musk said during a remote video interview on Feb.15. "I'm guessing probably towards the end of this year should be a good timing to find someone else to run the company, because I think it should be in a stable condition around, you know the end of this year." 

But in the end, the wait and the search went faster than expected.

"Excited to announce that I’ve a new CEO for X/Twitter," the tech tycoon said on May 11 in a tweet. "She will be starting in ~6 weeks!"

The billionaire, who added that he will transition to a more technical role, didn't provide the name of the new chief executive officer at the time.

"My role will transition to being exec chair & CTO, overseeing product, software & sysops," he said.

But 24 hours after crazy rumors, the billionaire has just given the name of the new CEO.

Linda Yaccarino

"I am excited to welcome Linda Yaccarino as the new CEO of Twitter!" Musk announced on May 12.

He added that she will focus "primarily on business operations, while I focus on product design & new technology."

"Looking forward to working with Linda to transform this platform into X, the everything app."

London-Based Seedcamp Launches $180 Million Fund to Back European Entrepreneurs

BY PYMNTS  |  MAY 17, 2023


Seedcamp has launched a $180 million fund to back European entrepreneurs.

The London-based seed-stage venture capital (VC) fund will use its Fund VI to lead investment and write first checks of up to $1 million in angel to seed rounds, Seedcamp said in a Wednesday (May 17) blog post.

“We remain sector-agnostic in our approach and are looking to back companies building foundational technologies spanning sectors, including the likes of artificial intelligence, cybersecurity, open source software, HealthTech, FinTech and more,” Seedcamp said in the post.

Together with the fund, Seedcamp has formed a new Seedcamp Expert Collective (SxC) that consists of more than 100 operators who will provide support to founders and teams via one-to-one sessions, workshops, content, micro-communities and, in some cases, angel investments, according to the post.

Seedcamp has been active for 15 years and now has a portfolio of more than 460 companies, including nine unicorns and two publicly listed companies, according to the post.

Its latest fund is almost double the size of its previous one, and the firm believes “now really is the moment to invest in the next decade of entrepreneurship,” Seedcamp said in the post.

“While volatility might be the current name of the game, at Seedcamp we are optimists at heart,” the firm said. “Over this time, we’ve also seen the European tech ecosystem mature, professionalize and join forces in times of crisis and experienced so many positive stories for Europe and across our Seedcamp Nation.”

This news comes on the same day that Finland-based VC firm Lifeline Ventures closed a fund at €150 million (about $163 million) and said it will use the fund to continue backing early-stage founders.

With this fund — which is its fifth and largest since its founding in 2009 — Lifeline Ventures will continue to focus on Finnish startups, for the most part, Sifted reported Wednesday.

EU early-stage startups can benefit from VC dry powder, Zeynep Yavuz, FinTech partner at early-stage VC firm General Catalyst, told PYMNTS in an interview posted in April.

“I believe it’s an amazing time to build a company,” Yavuz said, “and because the funding market has slowed down in the growth and late-stage space, a lot of the capital is shifting to early-stage.”

Snowflake in Talks to Buy Search Startup Neeva in AI Push

By Kevin McLaughlin, Jon Victor and Amir Efrati

May 17, 2023 1:36 PM PDT ·

Sridhar Ramaswamy. Photo by Getty
Sridhar Ramaswamy. Photo by Getty

Database software provider Snowflake has been in advanced talks to acquire Neeva, a search startup founded by former top Google ad tech executive Sridhar Ramaswamy, according to a person with direct knowledge of the discussions. Buying Neeva could help Snowflake offer artificial intelligence software that helps companies search for information in internal documents and data, according to people who do business with Snowflake.

Neeva primarily sells an ad-free web-search app for consumers, but it developed software that combines search with large-language models, which are trained on text to understand the nuances of speech and writing. That could fit with Snowflake’s efforts to help cloud customers use the kind of AI popularized by chatbots like ChatGPT that respond to conversational commands and can automate some business tasks. Snowflake is trying to catch up to rivals such as Microsoft’s Azure and Google Cloud that already sell access to such AI software.

• Snowflake aims to raise profile in AI
• Deal illustrates emergence of new kind of database
• Neeva’s search service has struggled

The deal shows how the emergence of ChatGPT is shaking up the enterprise software market, prompting companies such as Snowflake to rethink their growth strategy.

Unicorns Squeezed Between Capital Crunch and Blocked Exits

Some unicorns may have to risk raising a flat or down round or explore debt options due to a lack of exit choices, says PitchBook.

Vincent Ryan May 15, 2023

Unicorns Squeezed Between Capital Crunch  and Blocked Exits

U.S.-based “unicorns” — privately held startups valued at $1 billion or more — could face a tough 2023 as venture investment falls and the options to give investors a liquidity event shrink.

Record amounts of money in the venture ecosystem from 2013 to 2023 drove the aggregate number of unicorns in the United States from 35 to 704 for a collective valuation of nearly $2.4 trillion, according to PitchBook Data’s latest institutional research. 

However, there’s a growing imbalance between capital supply and demand, partly due to nontraditional venture investors fleeing the market. That will make it harder for unicorns to stay fully funded in the private markets while they make plans for an exit.

At the same time, the M&A and initial public offering markets are near a standstill, meaning VC-backed companies have a shortage of exit opportunities for founders, employees, and investors. 

This liquidity crunch has consequently trapped enormous amounts of value within highly valued organizations. — Vincent Harrison, PitchBook Data

Since 2022, just $78.7 billion in exit value has been generated within the U.S. VC market — an 87.2% decline from the record $768.3 billion achieved in 2021, according to PitchBook.

In a research commentary, “The Decline of Unicorn Acquisitions in a Conservative M&A Market,” PitchBook analyst Vincent Harrison said, “This liquidity crunch has consequently trapped enormous amounts of value within highly valued organizations.”

According to Pitchbook, there were only five acquisitions of unicorns in 2022, compared with 24 in 2021 and 17 in 2020.

Overall, only $39.6 billion in U.S. acquisition value has occurred since the beginning of 2022, “which makes it the least active year since 2015,” according to PitchBook. 

Dampened M&A

While the lack of IPOs is well-discussed, M&A deals are rare among unicorns. There are few companies with the excess cash to pursue purchases of unicorns, and those that do have “may be unwilling to spend that money on individual acquisitions, especially in the current market environment,” said Harrison.

Less than 25% of U.S.-based unicorns are profitable, said the PitchBook analyst. “While they often have high valuations and large amounts of funding, they also tend to have high burn rates as they invest heavily in market share and a ‘growth-at-all-costs’ model.”

In addition, the Biden administration’s antitrust crackdown makes “active tech acquirers such as Amazon, Apple, Google, and Meta — all of which are currently being investigated by the FTC or DOJ — uncertain whether their proposed deals will be approved or face lengthy antitrust reviews.”

The acquisition rate for unicorns is likely to remain limited for at least the remainder of 2023, reducing the number of liquidity options for these companies and “potentially reducing their bargaining power [or] negatively affecting valuations,” according to Harrison.

These relatively mature startups “will likely have to risk raising a flat or down round, explore debt options, or, in an ideal case, achieve positive cash flow to wait out a hostile financing environment,” according to PitchBook’s first quarter VC Valuations report.

Being labeled a “unicorn” has benefited some of these companies, allowing some to attract “additional pockets of venture funding from both traditional and nontraditional sources.”

With fewer investments from corporate venture capital, private equity firms, asset managers, and sovereign wealth funds, only $7.8 billion of late-stage deal value in the first quarter involved a nontraditional investor, according to PitchBook, about 10% of the sum invested in 2022.

The total number of late-stage deals completed in the first quarter of 2023 was slightly higher than the previous quarter. Still, the median deal size fell about 25% to $6 million, the lowest figure observed since the second quarter of 2017. (See the table below for a list of 2023 transactions so far.)

The subsequent capital crunch has substantially impacted valuations: The median late-stage valuation in Q1 fell to $5 million, an 8.3% decline from Q4. “The decrease in deal sizing speaks to the dearth of capital at the late stage,” according to Harrison.

PitchBook’s VC Valuations report estimates the capital-demand-to-supply ratio at the late stage to be 3.24x. That translates into $3.24 of capital sought by late-stage startups for every $1 of capital supplied by investors.

Startup of the Week

AI in your pocket: ChatGPT officially comes to iPhone with new app

App brings popular AI assistant to an official mobile client app for the first time.

BENJ EDWARDS - 5/18/2023, 10:48 AM

Screenshots of the ChatGPT iPhone app provided by OpenAI.
Enlarge / Screenshots of the ChatGPT iPhone app provided by OpenAI.

On Thursday, OpenAI released a free ChatGPT app for iPhone in the US that includes voice input support through its Whisper AI speech recognition model; it can also synchronize chat history with the web version of the AI assistant. The move brings ChatGPT to an official mobile client app for the first time.

ChatGPT, which launched in November, is an AI language model tuned for conversational input. Since then, it has expanded into a versatile AI assistant that can aid with tasks such as idea generation, compositional help, note summarization, personalized advice on various topics, and formatting or processing of text. It can also potentially serve as an educational resource if you trust the accuracy of its answers, which are sometimes inaccurate (we recommend double-checking anything it tells you).

Like on the ChatGPT website, users must log in to the ChatGPT app with an OpenAI account to use it, and the AI processing takes place off the device on OpenAI's servers, so it requires an Internet connection.

ChatGPT Plus subscribers have access to similar features as the web version, such as GPT-4, promises of "early access" to new features, and faster response times. In our tests we did not see beta access for ChatGPT with Browsing or ChatGPT Plugins.

In our early experiments with the new app, we found a bare-bones but functional application that serves as a much better interface to ChatGPT than attempting to use the ChatGPT website through a mobile browser. Early tests of the Whisper-based voice recognition proved buggy (often returning errors), perhaps due to overloaded servers on OpenAI's part, but those issues might be resolved soon.

OpenAI has started the app's rollout in the US, with plans to expand to additional countries in the coming weeks. "We’re eager to see how you use the app. As we gather user feedback, we’re committed to continuous feature and safety improvements for ChatGPT," writes OpenAI in an announcement blog post. According to OpenAI, "The release of the ChatGPT app for iOS is a step towards our goal of converting state-of-the-art research into practical tools, while continually increasing their accessibility."

Tweet of the Week