My teammates and I at
@OpenAI
are hiring ML engineers for research on LLM multi-step reasoning! We recently hit a SOTA 78% on MATH: . Our new plans are even more ambitious. If you want to pursue this with us send a resume/linkedin to michelle.kim
@openai
.com
@vgr
China: C- (the least warning so the hardest)
Korea: A (the church event a hard test, doing amazing)
Singapore: A, super impressive
Taiwan: A+, amazing.
Germany, France: D- , italy t -7
Italy, Iran: F (if the ICU is full before you lockdown, you too slow)
US: F (italy t-14)
copilot makes programming so much nicer, it is incredible to me that it doesn't have total adoption in the professionals who program as one of their contributions.
MEV is necessary for permisionless protocols. Designers chose the structure of the ordering protocol so that miners seeking to maximally extract value lead to the fulfillment of the design goals of the protocol.
This reminds me; if you know people who are into making the world better with computers and who could use some months in Berlin, let me know a spare room and board provided.
@halvorz
@SamoBurja
IMO, scenes and social networks are the real engine of amazing stuff in the world
so use it to support some kinda scene
even if that's just "my most creative friends"
give them a spare room rent-free for a while
host small backyard dinners so they can meet each other
AI doomerism boils down to belief that P(AGI kills us) > P(AGI helps us not get ourselves killed because it would also take it out). The fact that something is smarter than humans doesn't necessitate it can do better by getting rid of them. It might not be possible to be much ...
@robinhanson
the main bottlenecks are not training people but oxygen concentrators and non-invasive respirators; there is some work on trying to build up capacity to DIY manufacture these. Without those tools not much you can do, no matter the training (once things are severe)
"The Case for Universal Cloth Mask Adoption & Policies to Increase the Supply of Medical Masks for Health Workers" by me
@judy_chevalier
@NAChristakis
@thehowie
E Kaplan, A Ko (Chair of Epidemiology
@yale
) S Vermund (Dean of School of Public Health
@yale
)
I have been doing this. The main limits is that telco location (single carrier triangulation) creates very big radius of potential contact (using real data we where getting about 35K people in potential contact a infected person in a large metro) (1/n)
What we especially need is for tech people to outline to epidemiologists the full extent of the surveillance network *already deployed* commercially in the United States, and find out what parts of it will help their work. I and many other nerds are ready to help communicate that
I've been thinking about what the welfare optimal distribution of stake in a PoS system is, it be fun to talk about this with more people, so here is a early rough sketch (thanks to
@sxysun1
,
@AnettRolikova
,
@barnabemonnot
, for helpful comments on this)
@_alice_evans
it is striking that the emergency room gender ratio is opposite to the suicide rate gender ratio, makes me think "want to kill themselves" and "ER visit from self inflicted injuries" are not quite good proxies for each other
Some thoughts on a AMM design that minimizes MEV by directly eliciting the starting prices at a block to match end of block prices, at the cost of not providing liquidity between blocks.
@michael_nielsen
external-objective-guided science (as opposed to basic science where the objectives are internally driven by the furthering of the understanding itself)
a favorite recent paper " an equilibrium notion we call performative stability. Performative stability implies that the predictions are calibrated not against past outcomes, but against the future outcomes that manifest from acting on the prediction"
2023 defi pitch tl;dr
Here's a fundamental problem tracing back to the fact that there's no such thing as a free lunch.
Here's how we obfuscate that problem to make it harder for you to see it.
@ZakAttack1880
@waitbutwhy
@benshapiro
I think if you want to catch every human you have to go back close to Mitochondrial Eve, so more like ~5000th cousin (150k years ago)
One of my favorite papers ( ) started a line of work that tries to formalize some related intuitions; it turns out some fairly natural market making strategies for binary options are equivalent to online regret minimization. (1/n)
@vgr
1. Alternative and non-traditional medicine is largely quacks and net harmful practices. Just like mainstream medicine. The fundamental cause of both is the same: misaligned incentives. Outside antibiotics, vaccination and public sanitation, little works, alternative or not.
AI alignment in "large" non-zero sum worlds can be about constructing interfaces between the in-sillico cognitive domain and the carbon-based life that make their interactions complementary. Attempting to impose bans via the state is one way of loosing control by driving...
A PhD student asked me to be on his committee, flattered and terrified at the same time. My PhD thesis has been accepted subject to corrections, which I was in the middle of doing when the request arrived. Academic grandundlehood awaits you
@mdreid
America: the land where you have to pay for your own infectious disease treatment during an epidemic, but if you survive selling your antibodies is illegal.
A potential anti-pattern I've seen in in a couple of papers/posts around permissionless coordination devices, that seems subtle and interesting: to rely in the design of the system on some group of users not to coordinate on profitable, unobservable, collusion.
This would be fantastic; the same way ZK tech catalyzed by crypto is expanding the available design space of practical computational mechanism designs, so would MPC & enclaves
(1/2)
last week at BIU, Dan Boneh gave a presentation on MEV and laid out a prediction:
the same way zero knowledge cryptography was catalyzed by blockchains over the last years, mpc & enclaves research will be catalyzed by the mev problem in the next years.
i agree :)
@michael_nielsen
the NSF could run it like the IRS/SEC whistleblower schemes; calculate a percentage of the funds that the uncovering plausibly benefited and give some share as fellowship funding
a service that stenographically embeds private keys into photo-prints you send friends and family. Social wallet recovery via, "hey can you take a pic of that pic I sent you"
A Survey of LLM Hallucinations & How to Reduce Them
1) Overview of metrics & methods
2) Research progress on:
-Visual-language generation
-Abstractive summarization
-Data-to-text generation
-Machine translation
-Dialogue generation
-Generative Q&A
Nice analogy. We are at the KT Boundary for information. Archivists should lock down the pre-2023 information world. What comes after is going to be… different.
@ByrneHobart
equivalent; a currency is ultimately a claim on the output of those who have to meet future obligations in that currency. Debt or taxes denominated in it are two forms of such obligations.
@francoisfleuret
@IEEESpectrum
I have zero love for the IEEE (a completely captured institution that has turned into a parasite of the host it was nominally serves) but the series also contains, and linsk to at the top very contrasting articles,
@bertcmiller
@phildaian
Be interesting to figure out the tradeoffs between those 3 objectives. Intuitively it seems very hard to do all three to the max. You guys have anything written up on these tradeoffs?
solving evm parallelization, sharding execution (a la narwhal/bullshark worker architecture) and disk i/o feel like they all point to same problem: knowing what is going to get touched and knowing where (which thread, worker etc.) to route each storage access
if you want me not to call you are doomer for demanding others slow down lines of research, then you need to at least give some argument of why the above inequality holds, if it is a religious belief, completely unaffected by evidence, then doomer seems perfectly approapiate.
Prediction: Before the end of 2023, very expressive and high quality generative singing voices will become available open source. It will be a deeply transformative technology for music and culture.
@zacharylipton
or said differently, if you have just published in a journal a paper that will go on to get > 3 thousand citations, your next manuscript there will probably be reviewed by the editor.
@DrSarahLi
@jeremyphoward
he is explicitly *not* advocating the use of cloth masks in *healthcare* workers. The *population at large* using home made masks, frees up the higher quality masks for the health care users. They both improve outcomes, the home made masks relatively less so.
The closer to the merge the less incentive compatible for miners to be honest with pow flashbots and not defect and take all the MEV searches reveal. I wonder when this backwards induction dynamic hits.
I think there’s an information theoretical boundary on how much a single agent will do in your life.
It’s hard to believe in a single super agent that runs your life.
Humans need different faces. Different voices.
I think we’ll first discover this from prompting agents
@IvanWerning
@judy_chevalier
@NAChristakis
@thehowie
@Yale
Micro evidence for whether masks prevent you from transmitting the virus is harder to come by -- no one will let you do a randomized experiment where sick people don't wear masks (which should itself indicate how strong the prior is about the effect!),
@willbradbio
@anderssandberg
worth noting the octave is not a universal human percetual phenomenon, there are cultures where doubling of frequency does not result in pitch equivalence
I never realized just how fragile tokenization can be when you're crafting LLM prompts!
Say your model was trained to summarize with "\n\nTLDR:" and you decide to include an extra space after ": " so that the space is excluded from the generated output: it's different tokens.
best-estimate covid-19 fatality rates (from the pre-print in post bellow ):
Age 0-9: 0.0094%
Age 10-19: 0.022%
Age 20-29: 0.091%
Age 30-39: 0.18%
Age 40-49: 0.4%
Age 50-59: 1.3%
Age 60-69: 4.6% (range 3.8-5.4)
Age 70-79: 9.8% (range 8.2-12)
Age 80+: 18% (range 14-22)
Post Edited: Coronavirus age-specific fatality ratio, estimated using Stan, and (attempting) to account for underreporting of cases and the time delay to death. Now with data and code.
The data from mobile apps that sell your location, essentially A-GPS is more accurate but the coverage is much spottier, still the contact radious is quite large, not very practical for contact tracing, but enough for modelling the co-habitation and co-working networks (2/n)
@joshu
@naval
the first wave of this in PoW chains was self-defeating (the low work fork can be attacked), perhaps a second wave which makes more sense might form in Proof of Stake systems, that differentiate on community/governance?
@ConorMcMenamin9
@llllvvuu
this very recent work with Bruno shows that this problem with commitments breaking mechanisms that look incentive compatible in anon settings is quite insidious, for example pro-rata games look false-name/sybil proof until you consider commitments