Mr Robot

[Matt Shumer @mattshumer_]

LLaMA 2.0 is coming soon.

Seems like it will be:
- commercially usable
- closer to closed models like GPT-4/Claude 2 in capabilities

Excited to see the landscape shift.

https://t.co/OiD4fLQUy0

[Cameron R. Wolfe, Ph.D. @cwolferesearch]

Recently proposed open-source language models have placed an emphasis upon inference speed. Such work has shown us that inference speed can be improved by up to 5X (or more) by making some changes to the decoder-only transformer architecture. Here are three examples that have appeared frequently in recent work…

1. Flash attention: Self-attention is an O(N^2) operation, where N is the input length. Many efficient attention implementations (e.g., Reformer or Performer) have O(N) complexity but achieve minimal wall clock speedup due to IO latency. Flash attention solves this problem by reformulating attention in an efficient and IO-aware manner.

2. Multi-query attention: Multi-query attention shares key and value vectors between all attention head in a layer instead of performing a separate projection for each head. Although this change does not make training any faster, it significantly improves the inference speed of the resulting LLM.

3. Parallel layers: Most transformer architectures use a “serial” structure in each block. Alternatively, we can perform self-attention and the feed-forward transformation in parallel, followed by a single layer norm operation. This formulation does not deteriorate performance and can improve inference speed by parallelizing major operations in a transformer layer.

The impact: All three of the above modifications are added to the architecture of the recently-proposed Falcon-40B model. As a result, Falcon-40B can perform inference 5X faster than GPT-3 and even requires less compute for pre-training compared to comparable models like PaLM-62B or Chinchilla.

[Replit ⠕ @Replit]

State of AI Development

- 34000% growth in AI projects
- @OpenAI dominance
- Rise of Open Source with @huggingface
- The growth of @langchain
- And more

Read more: https://t.co/jmfuhIaQCT

[Matt Levine @matt_levine]

If this newsletter becomes the backbone of global value transfer, I expect the demand for Money Stuff to be considerable.
https://t.co/oSiLMc5mIr

[CoinDesk @CoinDesk]

Google will allow NFTs in Google Play store apps and @Starbucks recruited @AkuDreams for its next at-bat with NFTs.

@tobyleah wraps up this week's Web3 news:

https://t.co/3x6W7tAoXE

[CoinDesk @CoinDesk]

Crypto exchange giant @binance has laid off over 1,000 workers recently, with a big chunk of those happening this week, a source told the WSJ.

@HeleneBraunn reports
https://t.co/ahkABDrZl4

[Scientific American @sciam]

A classic test of self-control can carry complex cultural biases | Analysis https://t.co/062Jo5zgaH

[nader dabit @dabit3]

having worked in the blockchain space for ~2.5 years I've formed some unpopular opinions that I'd like to share.

1. the problems with security are laughably underestimated by almost everyone here.

most people act as if losing your identity and all of your money because you clicked the wrong button is not a massively huge, unacceptable, gaping problem. this still happens all the time, even to "veterans". this is such a problem that I find it tough to recommend to people who are not tech savvy to even try using most of this stuff.

2. the large majority of companies are absurdly overvalued.

at peak euphoria there was software that barely worked and with hardly any actual users valued at multiple billions of $, there are still protocols valued a billion $ or more who have yet to launch a working product and if launched, the majority of these projects will no where even close to live up to their marketing or self-imposed hype and often have critical issues they have yet to actually figure out that are part of what they promised, and are often just overly complex ponzi schemes who have successfully overpromised and confused investors enough to give them money.

3. I love most of the people I've met and have worked with.

that being said, there is a massive variance of people working in this space, more-so than anywhere else I've worked and by a long shot. this means being bombarded by a countless stream of grifters and people who want to simply use you as a financial instrument or a stepladder of some sort. when you reach a somewhat high profile this becomes unmanageable and will affect your mental health.

my dms and many of my tweets are bombarded by airdrop farmers or scammers. people will go to almost any length to get your attention only to often just be trying to get something out of you once you actually talk with them.

4. there are few use cases that actually make sense today.

finance / payments, digital ownership, data archival, recently our infrastructure is beginning to be able to handle the scale of things like gaming and social.

but as a developer building an app outside of these domains, it's extremely tough to find a use case that's valuable to the *average* person in the world, especially with the limitations of wallet UX and security.

instead everyone keeps building for the exact same set of users over and over. these are *extremely* talented builders, this isn't their fault, it just isn't possible today with the current state of the technology.

5. almost everyone you meet is founding a company of some sort, often competing with a hundred other similar products that do or are hoping to do the same thing, instead of collaborating. almost all with the hopes of launching a token and striking it rich.

yes, this happens in the traditional software world, but it is at an extreme in the blockchain world. innovation is great though, so this could be looked at as a good thing if you take into consideration that a percentage of these people are actual geniuses with good intentions who deserve to be successful, which is the case IME.

6. business models are fucked.

there are very few protocols or apps generating actual revenue, most are just hoping to launch a token. this is because it is very tough to build a product, get users, and monetize with the current state of the technology. there are a few outliers of course, but they are just that.

7. hype as a measure of integrity. it's all a blur.

the more hype a team, company, person, etc... tries to gather around themselves the less I can trust them. every day there are countless announcements all claiming to be revolutionary or game changing. there are teams who have raised hundreds of millions simply from tweeting out the right things.

instead the teams and people I've learned to trust are the ones who ship then shout, and who are pragmatic about their updates. @paradigm is a great example of this, only talking and gaining excitement once a big thing is ready, and usually it is of very high quality.

an announcement of an announcement or an announcement of a pre-announcement is not unheard of in this space and is a huge red flag.

>> conclusion
I have more I could go into but I'll stop for now.

the point of this isn't to be negative, I am one of the biggest believers of this technology.

instead it's to be a realistic view of how things are in hopes that we can improve the entire space.

again I will underscore that most of the people I've actually had the chance to work with, meet at conferences, or meet through acquaintances have been some of the most exceptionally smart, kind, and socially conscious people i've ever met.

even with all of the negatives, things have come a long way in the 2.5 years that I've been here. some of the issues i've outlined are things that we're working on at @AaveAave and @LensProtocol which is fulfilling work.

[ali @alitaylor]

this says all you need to know about apple vs meta

[The Economist @TheEconomist]

Sexual dysfunction lies beneath the same blanket of taboo and embarrassment as mental health once did. But tackling it can be simple and cheap https://t.co/UzhQq0TveL

[Lisa Abramowicz @lisaabramowicz1]

JPMorgan now has more than 13% of the nation’s deposits and 21% of all credit-card spending. https://t.co/Rx8PB79LxY

[Jonathan Weil @JonathanWeil]

KPMG gave Silicon Valley Bank a clean audit report 14 days before it collapsed. It blessed Signature Bank’s books 11 days before. Did they fail overnight or was it a long time coming? Like regulators, KPMG has many questions to answer. $SIVB $SBNY #KPMG https://t.co/bKQ1t6EPvQ

[Haseeb ＞|＜ @hosseeb]

Is the banking panic in the US extending to other countries? Have seen limited reporting on this.

[Chris Hayes @chrislhayes]

These billionaire VC's started a bank run on their own bank and then *immediately* pivoted to screaming for a government backstop. Incredible stuff.

[Mr Robot @russianbotty]

@carlquintanilla @carlquintanilla God bless you sir. True reporter.

[Carl Quintanilla @carlquintanilla]

August 2021.
Today.

[Carl Quintanilla @carlquintanilla]

JPMORGAN: SVB “was in a league of its own,” with an “unusually high reliance on corporate/VC funding .. and very low reliance on stickier retail deposits .. Bottom line: $SIVB carved out a distinct and riskier niche .. setting itself up for large potential capital shortfalls ..”

[Mr Robot @russianbotty]

@toly twitter mobs are noisy but I'm sure the back channeling is much more effective and helps in this like these.

[toly 🇺🇸 @toly]

I doubt the twitter mob had much influence on events. My bet is that the folks at the fed expected some banks to fail. The failure will accelerate moving assets into t bills and tighten credit which will slow the economy.

[Mr Robot @russianbotty]

Hmm. Seems interesting tech was so vocal could they have forecasted this? 🤔 I'd bet they gamed this out. If this happens we'll done techies. #NLCGW Never let a crisis good to waste.

[Carl Quintanilla @carlquintanilla]

Wow.

GOLDMAN: “In light of recent stress in the banking system, we no longer expect the #FOMC to deliver a rate hike at its March 22 meeting with considerable uncertainty about the path beyond March.” [Hatzius]

[Andrew Ross Sorkin @andrewrsorkin]

This. Is. Everything.

If everyone had this kind of financial literacy, the world would be better off.

Kudos to @Kyng_Kyren and family.

Volume on 🔊