throwaway9274

Yes, that’s correct. Good correction.

Unquantized model is here: https://huggingface.co/152334H/miqu-1-70b-sf

This strikes me as less a leak and more clever marketing from Mistral.

Sure, the eventual societal equilibrium of a world with instant AI-generated nude photography is that value of shame over nude photography inflates away to zero.

But that’s cold comfort to a teenage girl being mocked from the next lunch table today.

Twenty years ago was 2003. There are dozens of examples of AI x-risk thought experiments dating back to 1960.

As far as I can tell, all he did was open a forum for people to write fanfic about these earlier ideas.

The same thing that happens with DAN jailbreaks of GPT-4.

Nothing.

The barrier between bad actors and bad acts was never a shopping list.

Start small and be consistent. A few years ago at over 300lbs I started by walking around the park across the street. About .5 miles. Every day I tried to add 1%.

After six months I added a light jog for a mile.

Six months after that I was well enough to start HIIT classes. Six months after that I was strong enough to start real weight training and distance running.

Three years later I am down ~80lbs, able to run 10 miles at an 8:00 pace, and bench 205.

These aren’t extreme numbers for someone who started off reasonably healthy. But if you’ve been sedentary a long time, the difference in how you move and feel will be shocking.

Starting off can pose real challenges to your sense of self-worth, because you feel pathetic when you see how small your initial efforts are compared to what you see online and in media. It’s important to shift your mindset toward self-comparison, being slightly better than yourself yesterday.

If you’re struggling because you have chosen other priorities over your body (e.g. coding, family) in your 20s, I promise you, you will get there with consistent effort and compounded effort over time.

What’s the clinical definition of “tight shoulders”?

Chronic contraction of the upper trapezius.

[I]s it shown to correlate significantly with inhalation problems?

There’s not much in the medical literature beyond noting that that upper trapezius is an accessory muscle in expanding the lungs. (MDs correct me if this is wrong.)

Physiotherapy research is less rigorous but more expansive on the point. See e.g. Kim et al.

BACKGROUND: Forward head posture (FHP) causes changes in the strengths and rigidities of cervical muscles.

OBJECTIVE: The aim of this study was to investigate correlations between FHP and respiratory functions and the muscle activities of respiratory accessory muscles in young adults in their 20s.

METHODS: A volunteer sample of 33 healthy young adults participated in this study. Craniovertebral angle (CVA), cranial rotational angle (CRA), vital capacity (VC), forced vital capacity (FVC), forced expiratory volume at 1 second (FEV1), peak expiratory flow (PEF), maximal voluntary ventilation (MVV), and sternocleidomastoid (SCM) and upper trapezius activity ratios were measured.

RESULTS: Significant positive correlations were found between CVA and VC, FVC, FEV1, PEF, and MVV, and a significant negative correlation was found between CVA and SCM activity ratio. Significant negative correlations existed between CRA and VC and FVC, and significant positive correlations between CRA and SCM and upper trapezius activity ratios.

I wish OpenAI embedded the ChatGPT version date from the interface in the shared chats. It would save a lot of speculation.

I am on the September 25 version, which says its cutoff date is January, 2022.

100% of people know the most common 1000 words. The remainder of those who know more words fall into a consistent curve across languages that follows Zipf’s law. This is different than “most people know 1000 words on average.”

I said “parts of the recommender system code.”

This is the kind of highly emotional reaction that’s not helpful.

Yes, I am quite familiar with building ML models, both training and building my own for which I’ve been paid large sums of money, and I’m here to tell you that you don’t know what you’re taking about.

There’s so much more information about an ML system than just the trained model that is important for understanding the effects of the system on a society, and its legal, ethical, and social ramifications.

Just seeing the type of RS being used, the ranking approach, and the information on SimClusters is enough for RAI folks to start to understand the ecosystem effects and how that can show up downstream in social effects.

https://blog.twitter.com/engineering/en_us/topics/open-sourc...

Yeah, that’s the only bad argument.

He was forced to due to ad revenue plummeting as a reaction to his own actions and policies.

That’s incorrect. The need to diversify away from ad revenue was a topic discussed with Jack prior to the acquisition. The rationale was that advertisers effective control content moderation policies due to the revenue which they provide.

This is true. Whether it’s bad on or not depends on your viewpoint and ideological position. What advertisers want for now, e.g. with regard to social policy, may be aligned with what you or I want now.

But there’s no guarantee that will be true in the future.

I’d argue there’s a reasonable middle ground between “a bloated team” and “just the die-hards.

Layoffs are painful, and they were handled poorly. But there can be no doubt the prior company was massively overstaffed.

If you’re going to cut, generally you want to cut deep to prevent future rounds. Arguably not increasing the size of the first layoffs led to the second, and more people could have been preserved in total.

There’s very little virtue in “middle ground” in this context.

Community notes existed before Musk, and their role is to dispute a claim or provide context. They don’t moderate content in any way.

It doesn’t moderate content in any way… except for placing large labels to “dispute a claim or provide context.”

Sure, if you very narrowly constraint content moderation to the Trust and Safety definition of removing content and administering bans it doesn’t.

Birdwatch existed, but the prominence of the feature and improved reliability of the feature weren’t launched until after acquisition.

It solved the main edge cases for content moderation by only displaying labels when moderators who disagreed sufficiently on other issues agreed on that particular label.

It has significantly impacted the disinformation at scale problem for the better.

After the CPUC voted to approve the self-driving expansion, the same people that tried to block self-driving cars there pulled two regulatory levers.

First, the California DMV, with result seen here.

The incident that triggered this event happened when a human driver, who is still at large, hit a pedestrian who flew into the path of the Cruise vehicle.

The car executed a maneuver to pull over for a safety stop, but failed to recognize the woman who was dragged some distance.

Second, they reached to the Federal agencies via Nancy Pelosi’s office, with primary focus being the NHTSA, the National Highway Traffic Safety Administration.

The request was to gather data, which given the friendly posture of the NHTSA at this time will likely result in the opening of an investigation.

Both of these moves have some bite. The California DMV can suspend licensure as happened here.

But they’re fairly straight-shooter civil service types in the end, and they’ll eventually clear the vehicles for use. You can see this in part because of the action against Cruise only, rather than Waymo.

Some may attribute this to Waymo’s greater political sway in California. But my experience of the products themselves is that Waymo has a higher quality self-driving system.

The NHTSA is a bit trickier to predict.

The NHTSA’s powers are broad in theory and as granted by statute. They could theoretically issue an order to remove a type of vehicle from the road for imminent threat to public safety.

However, their typical modus operandi is to issue recalls and work with the auto companies. They won’t want to test these powers under the current court, so my opinion is that they won’t take much action.

Generally I appreciate Elon Musk’s work a lot. He created two world-changing companies in SpaceX and Tesla. I wanted him to be good at running Twitter/X.

It seemed like a good match. Staid company meets indefatigable executive.

His early moves prompted a lot of pushback but were generally wise: Diversified away from ad revenue. Introduced breaking changes to accustom the user base to faster development. Cut costs they couldn’t afford by shuttering a data center. Shrunk a bloated team to die-hards via whaling-and-culling. Not least, open sourced some recommender system code, and largely solved the content moderation problem via Community Notes.

Individually, all good steps. Sure, there were many misfires along the way: the verification system, launch of Twitter Blue, and erratic public ideation of potential features.

But I am ready to declare the Musk&Twitter/X experiment a failure.

The crux of the matter is this: Twitter exists as part of an ecosystem. Elon has alienated a large part of that ecosystem, both in terms of creators and advertisers.

I don’t know whether the network effect will tip or the business will run out of cash first, but I am confident it will not grow enough to justify the investment.

Maybe I’m wrong. Maybe alternate revenue streams and the value of Twitter’s data for AI training will be enough to keep it viable.

But I’m just not seeing the path.

Hopefully the ZIRP idea that dies is that VC and Big Tech can create companies off a conveyer belt.

VCs today go after shiny baubles, like “27 innovators under 27” nonsense, and have forgotten they’re mainly looking for high talent very resilient technical weirdos and any other characteristics are bonus.

The really talented technical people I know have what I call “eff you skills.”

They literally walk away over minor inconveniences like “you can’t use your own mouse here” or “you have to use VS Code and not your Vim setup.” Don’t like something management does? “Eff you.”

Last place I was at did one day in the office, and they all left without new jobs.

All are now employed again, most a level up.

The problem with this idea is that the highest performers are the group most likely to quit.

My guess is they’ll keep expanding it to higher groups as they feel out pushback.

It’s definitely more efficient for the airlines. Is it better for the customer? Depends on the segment.

Most budget customers respond only to price and will tend to prefer a lower price even with fewer amenities.

Everyone’s definition of a successful project varies, but 5k GitHub stars (at time of writing before people started gaming the metric) qualifies to me.

Modesty aside, the general principle reminds me of Alex Graveley, the eng lead on GitHub Copilot, saying his compensation for creating Copilot was only a $20k bonus and a title bump.

They’re both good illustrations that “make something people want” is a necessary step to turning code into wealth, but not a sufficient step.

Value capture != value creation.

A lesson every engineer learns with their first very successful project.

Personally I agree. My comment is about public opinion. I’m not arguing the public ought to feel this way but rather that they do feel this way.

That matters for thinking about what products will receive public push-back.

People don’t care about public recording as much as they did at the time of the glass launch.

The ubiquity of the smartphone camera and the public revelations about government surveillance did much to normalize recording in public.

It would be nice to establish norms and laws around anonymization of facial recordings and the need to auto-blur.

But I wouldn’t count on public outrage to hold the line on recording. I don’t anticipate the degree of public backlash we saw with glass.

This is not optimistic about advancing technology for the purposes of bettering the human condition. So the title is a bit misleading.

Words have meanings and we should adhere to them so communication does not break down into rhetoric.

It focuses on the social circumstances, or “human systems” for which it recommends re-engineering human societies to achieve a variety of goals.

It advocates for non-market economic systems, of which the only currently extant examples we have are socialist.

It advocates for slowing technological development to ease unspecified existential risks.

It offers that engineers can essentially make an inevitable societal decline a little bit less bad.

As much as Marc’s essay was a caricature of a considered techno-optimist position, this is the caricature of the inverse.

The problem with Marc’s exposition is that it mixes up techno-optimist ideas with personal right wing ideologies.

More fission reactors? Yes. Massively accelerate technological progress, and reduce the barriers to doing so? Yes. Is AI “extinction risk” overblown to the extent of being an obstacle to progress? Yes.

But markets aren’t self-balancing and do not prevent monopolies or cartels on their own. “Experts” generally know a lot, even though there’s gate keeping and credentialism. The ESG paragraph probably should just focus on the fact that making policy through finance is anti-democratic.

Beff Jezos, who Marc is cribbing here (and who is partially cribbing Nick Land) is pretty openly to the right.

In a document that’s being framed as capturing all of techno-optimism, these arguments are out of place. Without addressing the validity of the arguments themselves, they should have sanded off the explicitly political edges to maximize effectiveness.

We do not live in a post-scarcity world.

Every day people get up, enforce the laws, execute the national defense, bury themselves up to their elbows in humans’ abdomens to heal their broken bodies, and pick your turnips.

The world is not self-executing.

Market mechanisms coordinate that activity by transmitting information through the price signal. In their absence, the material abundance you mistake for post-scarcity would quickly collapse.

Both made bad decisions and should lose their money and be mocked by midwit redditors.

The difference is FTX was criminal negligence, SVB ordinary negligence. SVB was due to bad bets the bank was legally allowed to make, FTX was due to bad bets made illegally with stolen money.

Plausible deniability.

The Dread Pirate Roberts ordered passports on his own illegal website.

Smart criminals do stupid stuff all the time.

It would not surprise me at all if this is SBF or another insider’s insurance plan in action.

European societies have more fractious populations and a far more elite-mediated society than the U.S.

As a result, their union of states is both more fractious, in that it is a confederation and not a true federal government, and more elite-mediated, in that its function depends more upon the consensus among national elites than among national populations.

Put another way, it is more important that Paris agrees with Berlin than that Bautzen agrees with Lille.

Is that anti-democratic or just a more republican (small r) form of democracy?

I am not sure. But it is different.

We show that, if model weights are released, safety fine-tuning does not effectively prevent model misuse. Consequently, we encourage Meta to reconsider their policy of publicly releasing their powerful models.

The actual technology in the paper is cool, the work is well-done, but the conclusion “Meta should reconsider releasing model weights” does not follow.

Meta released the Llama 2 base model without the safety tuning already. It’s on HuggingFace, and chat finetunes of it based on uncensored datasets are popular.

So far no additional safety impacts are clear to me beyond the same issues caused by the availability of OpenAI’s APIs.

I expect the lack of major safety impacts will continue to be the case for two reasons.

First, non-existential risk concerns that do not implicate runaway AI such as “following harmful instructions” to create spam or explosives are a much higher barrier to entry with a smaller on-prem model than a clever prompt-based jailbreak of GPT-4.

Llama v2 could tell you how to build a biolab, but it would likely be wrong. And to do so you’d need to stand up your own hosting, get a dataset, LoRA the model, and then ask your evil question. Contrast that with copy/pasting the latest clever DAN jailbreak prompt into GPT-4.

Second, for x-risk concerns, the on-prem models are fundamentally not frontier models, which push beyond the performance of GPT-4. By definition open source hobbyists do not and will never have the resources to run or finetune frontier models. So any alignment work / x-risk testing can still take place prior to release of the model weights.

I am as concerned about AI risk as anyone, but the focus on open source LLMs seems like a distraction from real risks of large models already deployed like adtech and recommender systems.