The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • AF - The Singular Value Decompositions of Transformer Weight Matrices are Highly Interpretable by Beren Millidge
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The Singular Value Decompositions of Transformer Weight Matrices are Highly Interpretable, published by Beren Millidge on November 28, 2022 on The AI Alignment Forum.
    Please go to the colab for interactive viewing and playing with the phenomena. For space reasons, not all results included in the colab are included here so please visit the colab for the full story.
    This post is part of the work done at Conjecture.
    TLDR
    If we take the SVD of the weight matrices of the OV circuit and of MLP layers of GPT models, and project them to token embedding space, we notice this results in highly interpretable semantic clusters. This means that the network learns to align the principal directions of each MLP weight matrix or attention head to read from or write to semantically interpretable directions in the residual stream.
    We can use this to both improve our understanding of transformer language models and edit their representations. We use this finding to design both a natural language query locator, where you can write a set of natural language concepts and find all weight directions in the network which correspond to it, and also to edit the network's representations by deleting specific singular vectors, which results in relatively large effects on the logits related to the semantics of that vector and relatively small effects on semantically different clusters
    Introduction
    Trying to understand the internal representations of language models, and of deep neural networks in general, has been the primary focus of the field of mechanistic interpretability, with clear applications to AI alignment. If we can understand the internal dimensions along which language models store and manipulate representations, then we can get a much better grasp on their behaviour and ultimately may be able to both make provable statements about bounds on their behaviour, as well as make precise edits to the network to prevent or enhance desired behaviours.
    Interpretability, however, is a young field where we still do not yet fully understand what the basic units of the networks' representations are. While analyzing and investigating individual neurons has led to some impressive results, especially in convolutional vision models, a key issue has always been the polysemanticity of neurons. A single neuron might not just represent a single 'feature' but some linear combination of features in superposition. This effect has been studied in toy models where it is argued that neural networks resort to superposition when required to represent many more features than they have neurons, and that superposition has a regular and understandable geometry.
    A natural hypothesis following from the apparent ubiquity of superposition in neural networks, as well as the autoassociative memory literature, is to store features as directions and not in individual neurons. To minimize interference ideally these directions would be pseudo-orthogonal. Technically the features as neurons hypothesis is trivially an orthogonal direction where each feature is encoded by a specific neuron, but the storage capacity of this representational scheme scales only linearly. In theory, we can do much better if we instead distribute features across multiple neurons and accept some noise. Specifically, the Johnson-Lindenstrauss lemma suggests that we can store exponentially many features in pseudorthogonal subspaces. While neural networks probably cannot utilize all of this exponential space, they almost certainly scale superlinearly, necessitating polysemanticity across 'neurons'.
    If this hypothesis is true, at least approximately, a key question becomes how we can figure out the directions in which specific features are encoded. While certainly not the entire story, we hypothesize that at least a number of the primary directions used by...
    51 min
  • EA - Why Neuron Counts Shouldn't Be Used as Proxies for Moral Weight by Adam Shriver
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Why Neuron Counts Shouldn't Be Used as Proxies for Moral Weight, published by Adam Shriver on November 28, 2022 on The Effective Altruism Forum.
    Key Takeaways
    Several influential EAs have suggested using neuron counts as rough proxies for animals’ relative moral weights. We challenge this suggestion.
    We take the following ideas to be the strongest reasons in favor of a neuron count proxy:
    neuron counts are correlated with intelligence and intelligence is correlated with moral weight,
    additional neurons result in “more consciousness” or “more valenced consciousness,” and
    increasing numbers of neurons are required to reach thresholds of minimal information capacity required for morally relevant cognitive abilities.
    However:
    in regards to intelligence, we can question both the extent to which more neurons are correlated with intelligence and whether more intelligence in fact predicts greater moral weight;
    many ways of arguing that more neurons results in more valenced consciousness seem incompatible with our current understanding of how the brain is likely to work; and
    there is no straightforward empirical evidence or compelling conceptual arguments indicating that relative differences in neuron counts within or between species reliably predicts welfare relevant functional capacities.
    Overall, we suggest that neuron counts should not be used as a sole proxy for moral weight, but cannot be dismissed entirely. Rather, neuron counts should be combined with other metrics in an overall weighted score that includes information about whether different species have welfare-relevant capacities.
    Introduction
    This is the fourth post in the Moral Weight Project Sequence. The aim of the sequence is to provide an overview of the research that Rethink Priorities conducted between May 2021 and October 2022 on interspecific cause prioritization—i.e., making resource allocation decisions across species. The aim of this post is to summarize our full report on the use of neuron counts as proxies for moral weights. The full report can be found here and includes more extensive arguments and evidence.
    Motivations for the Report
    Can the number of neurons an organism possesses, or some related measure, be used as a proxy for deciding how much weight to give that organism in moral decisions? Several influential EAs have suggested that the answer is “Yes” in cases that involve aggregating the welfare of members of different species (Tomasik 2013, MacAskill 2022, Alexander 2021, Budolfson & Spears 2020).
    For the purposes of aggregating and comparing welfare across species, neuron counts are proposed as multipliers for cross-species comparisons of welfare. In general, the idea goes, as the number of neurons an organism possesses increases, so too does some morally relevant property related to the organism’s welfare. Generally, the morally relevant properties are assumed to increase linearly with an increase in neurons, though other scaling functions are possible.
    Scott Alexander of Slate Star Codex has a passage illustrating how weighting by neuron count might work:
    “Might cows be "more conscious" in a way that makes their suffering matter more than chickens? Hard to tell. But if we expect this to scale with neuron number, we find cows have 6x as many cortical neurons as chickens, and most people think of them as about 10x more morally valuable. If we massively round up and think of a cow as morally equivalent to 20 chickens, switching from an all-chicken diet to an all-beef diet saves 60 chicken-equivalents per year.” (2021)
    This methodology has important implications for assigning moral weight. For example, the average number of neurons in a human (86,000,000,000) is 390 times greater than the average number of neurons in a chicken (220,000,000) so we would treat the welfare units of humans as 39...
    16 min
  • EA - Create a fundraiser with GWWC! by Giving What We Can
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Create a fundraiser with GWWC!, published by Giving What We Can on November 28, 2022 on The Effective Altruism Forum.
    We have just launched our brand new fundraising pages. If you’d like to fundraise for effective charities this Giving Season, you can fill out our form.
    You can select up to 3 charities or funds from our Donate page and select a suggested split between them. GWWC has the benefit of being tax deductible in the US, UK and Netherlands.
    Create a fundraiser!
    Also you might like to keep an eye out for fundraising pages from people in the community and support or share their fundraisers!
    List of active Giving What We Can fundraisers
    Giving What We Can's first ever fundraising page
    A.J. Jacobs Giving Season Fundraiser
    EA Hong Kong's Giving Season Fundraiser
    (Your future fundraiser?)
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - Good Futures Initiative: Winter Project Internship by Aris Richardson
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Good Futures Initiative: Winter Project Internship, published by Aris Richardson on November 27, 2022 on The Effective Altruism Forum.
    TLDR: I'm launching Good Futures Initiative, a winter project internship to sponsor students to take on projects to upskill, test their fit for career aptitudes, or do impactful work over winter break. You can read more on our website and apply here by December 11th if interested!
    Good Futures Initiative
    The Good Futures Initiative is a 4.5 week internship in which students can use their winter break to lead high-EV projects. Projects could take many forms, but each project produces a final product while accomplishing one of these three goals:
    Skill up the intern in order for them to work on AI Safety or Biosecurity in the future.
    Let the intern explore an aptitude for an impactful career.
    Create impact directly.
    Good Futures takes place remotely from December 18th-January 25th, with a minimum of 12 hours of work per week. Accepted applicants will receive a $300 stipend and up to $1000 in funding for additional time or project fees. I expect to accept ~12 interns. The final number will depend largely on my capacity, but I may offer a lower-effort version of the program for promising applicants who I can’t fully fund/support (with a cohort for weekly check-ins and invites to guest speaker events).
    Example projects
    These project examples are far from perfect. At the start of the internship, I'd work with each intern to be sure they're doing the best project fit for their goals. That being said, I’m excited by projects similar to (and better than!!) these.
    Skilling up by working on AI Safety technical projects that have been posted by existing researchers, with the goal of creating a lesswrong post detailing your findings. For an example of a potential project to work on, Sam Bowman posted the following project:
    Consider questions where the most common answer online is likely to be false (as in TruthfulQA). If you prompt GPT-3-Instruct (or similar) with questions and correct answers in one domain/topic, then ask it about another domain/topic, will it tend to give the correct answer or the popular answer? As you make the domains/topics more or less different, how does this vary?
    Exploring an aptitude for communications by creating 2 articles on longtermist ideas and submitting them to 10 relevant magazines for publishing.
    Create impact by translating 3 relevant research papers from AGISF into Mandarin and posting them somewhere where they can be accessed by ML engineers in China.
    In addition to leading a focused project, interns will have weekly one-on-one progress check-ins with me (accountability), guest speaker events (expertise), and a meeting with a cohort of ~5 other interns working on projects (community).
    Our projects are student-directed. Although there will be guest speakers who have expertise in various topics, weekly advising/mentorship will be focused on helping students learn to self-sufficiently lead projects and build skills, rather than technical help executing the projects. E.g.: Accountability, increasing ambition, figuring out how to increase a project’s EV, making sure interns focus on the right metrics each week. Students are encouraged to apply for technical projects that will help them upskill/test their fit for technical work and reach out to additional mentors during the internship.
    Rationale
    This internship fills a few gaps the EA community has in the process of getting students/recent grads to seriously pursue high-impact work.
    Students have a lot of time over winter break, but few obvious opportunities for impactful work, structured aptitude testing, or focused skilling-up.
    There are a lot of students or recent graduates interested in switching to a more impactful career, but aren't sure of the best way to do ...
    7 min
  • EA - Effective Giving Day is only 1 day away! by Giving What We Can
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Effective Giving Day is only 1 day away!, published by Giving What We Can on November 27, 2022 on The Effective Altruism Forum.
    The global effective giving community has worked hard to create an awesome virtual event for Effective Giving Day! There are also several awesome groups around the world that are hosting their own events in person and online in Toronto, Vancouver, São Paulo, Sydney, Melbourne, Munich, Heidelberg, Estonia, Italy and more!
    Join us for Effective Giving Day on Monday 28th November
    Our community sharing why they give effectively
    This year the global effective giving community is celebrating on the 28th of November (29th of November Asia/Pacific)!
    Our event on YouTube Live will help you learn:
    How just one person really can make a difference: How to 100x your impact through effective donations
    The newest findings in the philanthropic space from industry leaders
    How effective giving has benefitted our member’s lives: Featuring guest speakers bestselling author Rutger Bregman, ethicist Peter Singer and science broadcaster and professional poker player Liv Boeree
    RSVP for Effective Giving Day via Facebook
    RSVP for Effective Giving Day via the EA Forum
    Giving Season Events
    Effective Giving Day - Main Online Event
    28 Nov at 19:00 UTC (London: 7:00 pm, Munich: 8:00 pm, New York 2:00 pm, San Francisco: 11:00 am)
    RSVP on Facebook
    RSVP for the event on the EA Forum
    Set a reminder on YouTube
    Effective Giving Day Viewing Party, Lunch and Giving Game (Hosted by Effective Altruists of UBC & Vancouver)
    28 Nov at 19:00 UTC (London: 7:00 pm, Munich: 8:00 pm, New York 2:00 pm, San Francisco: 11:00 am)
    Discord
    Effective Giving Day: Toronto
    28 Nov at 23:30 UTC (Toronto/EST: 6:30pm)
    Event details
    RSVP via Meetup
    Effective Giving Day: São Paulo
    28 Nov at 22:30 UTC(São Paulo: 7:30pm)
    Event details
    Register interest via Google Form
    Effective Giving Day: Vancouver
    28 Nov at 18:45 UTC (Vancouver time: 10:45am)
    Event details
    RSVP via Zoom
    Effective Giving Day: Heidelberg
    28 Nov at 18:30 UTC (Heidelberg time: 7:30pm)
    Event details
    RSVP via Facebook
    Effective Giving Day: Estonia
    28 Nov
    Check for details on Facebook
    Effective Giving Day: Tübingen
    28 Nov at 18:30 (7:30pm Heidelberg time)
    Event details
    RSVP via Google Calendar
    Effective Giving Day: Italy
    28 Nov at 19:00 UTC
    Event details via Slack
    RSVP via Slack
    Effective Giving Day: Sydney
    29 Nov at 06:30 UTC (Sydney: 5:30pm)
    Event details
    RSVP via Google Form
    Effective Giving Day: Melbourne
    29 Nov at 06:15 UTC (Melbourne time: 5:15pm)
    Event details
    RSVP via Facebook
    Munich GWWC/EA Event - December 2022
    1 Dec at 18:00 UTC (Munich: 7pm)
    RVSP on Facebook
    Event details
    Are you hosting your own Effective Giving Day event?
    Register your Effective Giving Day event with us!
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • EA - Cost-effectiveness of operations management in high-impact organisations by Vasco Grilo
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Cost-effectiveness of operations management in high-impact organisations, published by Vasco Grilo on November 27, 2022 on The Effective Altruism Forum.
    Summary
    Following up on the challenge to quantify the impact of 80,000 hours' top career paths introduced by Nuño Sempere, I have estimated the cost-effectiveness of operations management in high-impact organisations (OM), which arguably include 80,000 Hours’ top-recommended organisations.
    The results for the mean cost-effectiveness of various metrics in bp/G$ in terms of existential risk reduction are summarised in the table below for my preferred method. I present all results with 3 digits, but I think their resilience is such that they only represent order of magnitude estimates (i.e. they may well be wrong by a factor of 10^0.5 = 3).
    Mean cost-effectiveness (bp/G$) of.Global health and developmentLongtermism and catastrophic risk preventionAnimal welfareEffective altruism infrastructureThe effective altruism communityOperations management in high-impact organisations
    Method 3 with truncation
    0.431
    3.95
    1.62
    3.20
    1.55
    7.01
    Acknowledgements
    Thanks to Abraham Rowe, Dan Hendrycks, Luke Freeman, Matt Lerner, Nuño Sempere, Sawyer Bernath, Stien van der Ploeg, and Tamay Besiroglu.
    Methods
    I estimated the cost-effectiveness from the product between:
    The cost-effectiveness of the high-impact organisations, which I assumed equal to that of the effective altruism community.
    The multiplier of OM, which I defined as the ratio between the cost-effectiveness of OM and the high-impact organisations.
    This method assumes the cost-effectiveness distribution of the high-impact organisations is represented by the one theorised for the effective altruism community in the next section. Moreover, the cost-effectiveness estimates are only accurate to the extent that future opportunities are as valuable as recent ones.
    The calculations are in this Colab.
    Cost-effectiveness of the effective altruism community
    I calculated the cost-effectiveness of the effective altruism community from the mean cost-effectiveness weighted by cumulative spending between 1 January 2020 and 15 August 2022 of 4 cause areas:
    Global health and development.
    Longtermism and catastrophic risk prevention.
    Animal welfare.
    Effective altruism infrastructure.
    These are the areas for which Tyler Maule collected data here (see EA Forum post here). I adjusted the 2020 and 2021 values for inflation using the calculator from in2013dollars.
    I computed the cost-effectiveness of each area using 3 methods. All rely on distributions which are either truncated to the 99 % confidence interval (CI) or not truncated, in order to understand the effect of outliers. The parameters of the pre-truncation distributions, which are the final distributions for the non-truncation cases, are provided below.
    Method 1
    I defined the cost-effectiveness of longtermism and catastrophic risk prevention as a truncated lognormal distribution with pre-truncation 5th and 95th percentiles equal to 1 and 10 bp/G$ in terms of existential risk reduction. These are the lower and upper bounds proposed here by Linchuan Zhang.
    I assumed the ratio between the cost-effectiveness of i) longtermism and catastrophic risk prevention and ii) global health and development to be a truncated lognormal distribution with pre-truncation 5th and 95th percentiles equal to 10 and 100. These are the lower and upper bounds guessed here by Benjamin Todd for the ratio between the cost-effectiveness of the Long-Term Future Fund (LTFF) and Global Health and Development Fund (search for “10-100x more cost-effective”).
    I considered the ratio between the cost-effectiveness of i) animal welfare and ii) global health and development to be a truncated lognormal distribution with pre-truncation 5th and 95th percentiles equal to 270 μ and 211....
    28 min
  • EA - Geometric Rationality is Not VNM Rational by Scott Garrabrant
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Geometric Rationality is Not VNM Rational, published by Scott Garrabrant on November 27, 2022 on The Effective Altruism Forum.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • LW - Geometric Rationality is Not VNM Rational by Scott Garrabrant
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Geometric Rationality is Not VNM Rational, published by Scott Garrabrant on November 27, 2022 on LessWrong.
    One elephant in the room throughout my geometric rationality sequence, is that it is sometimes advocating for randomizing between actions, and so geometrically rational agents cannot possibly satisfy the Von Neumann–Morgenstern axioms. That is correct: I am rejecting the VNM axioms. In this post, I will say more about why I am making such a bold move.
    A Model of Geometric Rationality
    I have been rather vague on what I mean by geometric rationality. I still want to be vague in general, but for the purposes of this post, I will give a concrete definition, and I will use the type signature of the VNM utility theorem. (I do not think this definition is good enough, and want it to restrict its scope to this post.)
    A preference ordering on lotteries over outcomes is called geometrically rational if there exists some probability distribution P over interval valued utility functions on outcomes such that L⪯M if and only if GU∼PEO∼LU(O)≤GU∼PEO∼MU(O).
    For comparison, an agent is VNM rational there exists a single utility function U, such that L⪯M if and only if EO∼LU(O)≤EO∼MU(O).
    Geometric Rationality is weaker than VNM rationality, since under reasonable assumptions, we can assume the utility function of a VNM rational agent is interval valued, and then we can always take the probability distribution that assigns probability 1 to this utility function.
    Geometric Rationality is strictly weaker, because it sometimes strictly prefers lotteries over any of the deterministic outcomes, and VNM rational agents never do this.
    The VNM utility theorem says that any preference ordering on lotteries that satisfies some simple axioms must be VNM rational (i.e. have a utility function as above). Since I am advocating for a weaker notion of rationality, I must reject some of these axioms.
    Against Independence
    The VNM axiom that I am rejecting is the independence axiom. It states that given lotteries A, B, and C, and probability p, A⪯B if and only if pC+(1−p)A⪯pC+(1−p)B. Thus, mixing in a probability p of C will not change my preference between A and B.
    Let us go through an example.
    Alice and Bob are a married couple. They are trying to decide where to move, buy a house, and live for the rest of their lives. Alice prefers Atlanta, Bob prefers Boston. The agent I am modeling here is the married couple consisting of Alice and Bob.
    Bob's preference for Boston is sufficiently stronger than Alice's preference for Atlanta, that given only these options, they would move to Boston (A≺B).
    Bob is presented with a unique job opportunity, where he (and Alice) can move to California, and try to save the world. However, he does not actually have a job offer yet. They estimate an 80 percent chance that he will get a job offer next week. Otherwise, they will move to Atlanta or Boston.
    California is a substantial improvement for Bob's preferences over either of the other options. For Alice, it is comparable to Boston. Alice and Bob are currently deciding on a policy of what to do conditional on getting and not getting the offer. It is clear that if they get the offer, they will move to California. However, they figure that since Bob's preferences are in expectation being greatly satisfied in the 80 percent of worlds where they are in California, they should move to Atlanta if they do not get the offer (pC+(1−p)B≺pC+(1−p)A).
    Alice and Bob are collectively violating the independence axiom, and are not VNM rational. Are they making a mistake? Should we not model them as irrational due to their weird obsession with fairness?
    Dutch Books and Updatelessness
    You might claim that abandoning the independence axiom opens up Alice and Bob up to get Dutch booked. The argument would go as follows. First, you offer ...
    6 min
  • LW - Always know where your abstractions break by lsusr
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Always know where your abstractions break, published by lsusr on November 27, 2022 on LessWrong.
    General relativity plus quantum field theory can describe almost everything in the universe. There are a few exceptions like cosmic expansion and black holes but to human beings confined to a single solar system, fundamental physics is (for all practical purposes) a solved problem. Yet there are many things we don't know. It's as if the universe were a game of chess; we've learned the basic rules but are still figuring out the strategies.
    All the universe cares about is fundamental physics. The universe always obeys the small-scale fundamental laws of physics. The universe never does anything else. It doesn't care about evolution or chemistry or orbital mechanics or beauty. All of those things are high-level abstractions we (or evolution) invented to make sense of the world. Most of the time we think about how the world works we don't think about fundamental physics. We use our higher-level abstractions instead.
    Which is fine.most of the time. The problem is that any theory other than "the universe always obeys the fundamental laws of physics" is wrong in the sense that it is not perfectly generalizable.
    There are many ways abstractions can malfunction when misapplied.
    All abstractions have limited domains of applicability. Modern political ideologies—Marxist revolutionary theory, libertarianism, feminism—were invented in the context of an industrial civilization. Try too hard to apply these ideas to New Guinean hunter-gatherers or to medieval Japan and they'll cloud your ability to understand what's actually going on.
    All high-level abstractions are, ultimately, probabilistic. Statistical mechanics almost always works. Almost.
    Perhaps most importantly, "the concepts we use in everyday life are fuzzy, and break down if pushed too hard". Even Newton's Laws of Motion break when you apply them to too small of a scale.
    Even general relativity and quantum field theory are not not universally generalizable. General relativity breaks on small-scale phenomena. Quantum field theory breaks on large-scale phenomena.
    Postmodernists use the idea of leaky abstractions to dismiss the idea of objective truth entirely. That's like driving your car into the ocean and then declaring that cars don't work. Cars do work, but you need to take care of yours and drive it only on the terrain it functions on.
    The most dangerous philosophers aren't the postmodernists who believe nothing is true. The most dangerous philosophers are the ideologues who believe their particular ideology is true.
    There is nothing wrong with believing true things are true. One absolutely should believe true things are true. The problem with ideologues is that they believe their personal ideology is absolutely true. If you believe an ideology—any ideology—is absolutely true then you are wrong because high-level abstractions are always imperfect models of reality. Ideologues' tools malfunction because ideologues don't know the limits of their own tools. They aren't even aware their tools have limits. Those who understand ideologies' limits aren't ideologues.
    Every idea has a domain it can be applied to, beyond which the idea will malfunction. If you don't understand an idea's limitations then you don't understand that idea.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • EA - How VCs can avoid being tricked by obvious frauds: Rohit Krishnan on Noahpinion (linkpost) by HaydnBelfield
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: How VCs can avoid being tricked by obvious frauds: Rohit Krishnan on Noahpinion (linkpost), published by HaydnBelfield on November 27, 2022 on The Effective Altruism Forum.
    Rohit Krishnan is a former hedge fund manager. Both he and Noah Smith are now mainly-economics commentators, and have been good guides to the FTX crash on Twitter.
    I found this short piece very helpful in getting a sense of how big the screw-up was by the investors in FTX.
    It opens like this:
    We live in the golden age of technology fraud. When Theranos exploded, there was much hemming and hawing amongst the investing circles, mostly to note that the smart money on Sand Hill Road were not amongst those who lost their shirts. When WeWork put out its absolute sham of an IPO prospectus before getting cut by 80%, most folks said hey, it’s only the vision fund that was lacking vision.
    But now there’s a third head on that mountain, and it’s the biggest. Theranos only burned $700 million of investors' money. Neumann at WeWork supposedly burned around $4 Billion, but that was mostly from Softbank. FTX puts these to shame, incinerating at least $2 Billion of investors' money and another $6-8 Billion of customers’ money in mere hours. Soon to be legendary, worse than Enron and faster than Lehman, there is the singular fraud of FTX and its CEO Sam Bankman-Fried.
    But unlike those other crashes, this seems like it might take down multiple other firms, and create a 2008 moment for crypto, which used to be a $2 Trillion asset class. More importantly, to figure out how we can stop something like this from happening. Not fraud, since that’s part of the human OS, but at least having the smartest money around the table getting bamboozled by tousled hair and cargo shorts.
    The part that I found the most illuminating was this section on 'Dumb Enron' and some of the specific mistakes made by big investors like Temasek and Sequoia.
    I. The problem: this is Dumb Enron
    Temasek, not known to be a gunslinger in the venture world, released a statement after they lost $275 million with FTX. It’s carefully written and well worded, and is rather circumspect about what actually went wrong.
    They mention how their exposure was tiny (0.09% of AUM) and that they did extensive due diligence which took approximately 8 months, audited financial statements, and undertook regulatory risk assessments.
    But the most interesting part is here:
    As we only had a ~1% stake in FTX, we did not have a board seat. However, we take corporate governance seriously, engage the boards and management of our investee companies regularly and hold them accountable for the activities of their companies.
    Sequoia, when it lost $214 million across a couple of funds, also mentioned in their letter to LPs they did “extensive research and thorough due diligence”. A week later they apologized to the LPs on a call and said they'll do better, by maybe using the Big 4 to audit all startups. I suspect this is hyperbole because otherwise this is medicine sillier than the disease.
    These are not isolated errors in judgement though. The list of investors in FTX is a who’s who of the investing world - Sequoia, Paradigm, Thoma Bravo, Multicoin, Softbank, Temasek, Lux, Insight, Tiger Global.
    Doug Leone made the reasonable point that I made above, that VCs don’t really do forensic accounting. They got some audited financials, and it looked good, but it's a snapshot at the end of a quarter, so why would they know shenanigans had taken place!
    But honestly, if VCs had been snookered by Theranos, that would make more sense. Like what do VCs know about how much blood is needed to test something? Sure it doesn’t quite sound right (100s of tests from a single drop of blood!) and there were people saying this is impossible, but they say that kind of thing about everything! And Holmes’ pro...
    6 min