1. 1- Most diseases have a variable course of evolution. This means that a test like this ends up being more of a "risk factor" analysis than predictive of disease - think 23AndMe type risk associations rather than BRCA1 risk associations.
2. 2- If the focus of the test is sensitivity, as is often the case in early detection, then the test will almost certainly result in overdiagnosis of the disease. Here's a good article on how that plays out with imaging findings of "incidentalomas" (i.e. tumors that would likely not have been treated had they not been discovered incidentally). With a condition as broad as dementia, it seems very likely to result in overdiagnosis.
3. 3- If there aren't any treatments that alter the course of the disease, then the finding basically makes the patient needlessly worried and can frequently lead to unnecessary testing and treatment.

In short, unless there's specific reason to believe that the test will actually result in better and more appropriate treatment (e.g. it indicates whether a specific treatment will or will not work), be wary of it.

It sometimes pays to listen to the kooks in the fringes, if you're firm enough on your own feet to not have to throw a hissy fit while listening to things you don't necessarily agree with. You can find little seeds of interesting possibility scattered throughout the lunatic fringe. And let's be honest, a scientist given to hyperbole to promote his findings is hardly the most egregious character in our current society. He's just a bit outside the standard deviation of what current science deems acceptable.

This is one of the best things I've read on the internets in a long time. I think it's because you are an alien here to bring us ancient transcendental wisdom about all our base. :-]

Seriously, it's a great point that often the kooks are open to ways of thinking that others are not. As long as you can expect a sensitivity of a couple percent in terms of seeing something meaningful instead of fantasizing, there can be a lot to learn.

As with all things AI (and actually for many technologies stretching into the past), there is a raft of stories along the lines of:

Can you believe what technology X can do?
It can help with hard problem Z!
Here is a study where it showeld a positive result!
We should definitely use technology X for everything!

In this case, as in so many, technology X is being pushed by some company that is selling it and funded the study. I've created, run, and published studies like that myself. They can have meaningful findings, but often the company's goal is derived from only the positive findings:

Translation:

What I want to know is where the studies are that show how often people who need mental health help try a technology interaction with or without AI and are turned off, discouraged, frustrated. But nobody likes to measure that, and anyone talking about AI is going to report stuff people love about it.

We'll have to wait 5 more years until people can say "AI's not the answer after all".

I don't disagree but I'd put it a little differently and perhaps a little more charitably.

From what I've seen drug candidates emerge when someone discovers a substance that modulates some biological process in a disease that has unsatisfactory treatment options. The key is that all along the path, people know what the clinical problem is and focus on developing a drug that makes treatment work better. From what I can tell neither Wojcicki - despite having done some time in biotech investing early in her career - nor the rest of the company has made it their business to really study any disease deeply.

23AndMe's concept was always much more tech-based: mine the medical literature for disease associations and then link those findings to the results of DNA testing and synopsize the relevant literature findings for people. It was founded back when Quantified Self was the future and was supposedly going to usher in a new era of health and well-being rather than just drowning us in meaningless and unmanageable data. The basic problems for 23AndMe are that DNA is not that predictive of how diseases evolve, the medical literature as a corpus is messy and incoherent, and "lifestyle hacks" are not a firm basis for preventing or managing disease. I don't see how they change the parameters of their business without trying to dig more deeply into some disease and work more closely with those specimens and data.

As far as their DNA database, it's a little hard to see how it is a great drug discovery tool. One thing that could be interesting is if the person could be linked to long-term outcome data (e.g. via a Flatiron Health type entity). That would let someone look at DNA at t=0 for a lot of people. Even so, because the specimen is just a saliva sample, it's hard to see how it tells that much of a story beyond what's known about hereditary disease.

From the EULAs I've seen, I'd stay the hell away.
Samsung makes clear that it will collect and use absolutely every bit of your data that it touches and sees no legal impediments to doing so.

That's not to say MS, Apple, and Google don't also try to snarf up your data. But at least they pay lip service to things like the California Data Privacy Act...

It is true that India and China has vastly increasing emissions. And your point about double-talk is very well taken. I am also frustrated and terrified by the lagging pace of progress on decarbonization.

However, I don't think India or China don't believe in the climate emergency - they just don't think their countries' respective publics will tolerate the consequences of not adding cheap power immediately. Sadly, I think that calculus is accurate - in terms of public perception and power, lots of people express the desire to do something about climate change, but they are either resistant to change or easily manipulated by oil and gas into opposing most concrete steps.

I believe India's calculus is quite different than China's. China is in a precarious position economically, and they are wary of any economic disruption from high energy prices. They are actually installing solar at the fastest pace in the world, but they are also adding a lot of coal at the same time. China has vast reserves of coal but nearly none of oil. They are moving rapidly to EVs, which allows them to substitute locally sourced electricity for imported oil. So they question is basically how fast they can build out solar and decommission coal.

India I know less about, but I think they see themselves as well-positioned to capitalize on China's demographic decline and do a lot of industrial production. Since they have an enormous young population seeking employment, I'd guess they are betting on cheap energy now and hoping that someone else comes up with sequestration methods they can use.

And in both places, I don't think a Xi or Modi can actually just mandate that all coal-fired plants in the pipeline not be built. They have a lot of power, but in both countries there is an administrative state that also has significant influence. It's much, much harder to take away public dollar goodies than to give them out.

At this point it's pretty much a giant bet on future carbon capture. Maybe the good Sultan will come up with a way to force the oil and gas industry to actually commit real resources to carbon capture, but I too am skeptical. More likely is that they'll do some greenwashing and hope for someone else to do something.

I'd go so far as to say that this is anticompetitive activity and deserves complaints to the FTC. Here is the link to make a complaint about unfair business practices:
https://reportfraud.ftc.gov/#/...

As developers, I think we have an important voice on this subject and encourage all to submit their thoughts to the FTC.

Whose CDN?

While the TLS part is good, as long as they retain the ability to associate URLs and other stuff in the original request headers, they are still positioned as a central source of info on browsing habits.
I am quite sure that they won't introduce so much privacy as to break existing features in Google Analytics or keyword buying.

Color me skeptical that this will not increase their control over data on browsing habits, no matter what the technical implementation is.

This is spot-on. There are many problems, but one of the worst is employer-sponsored insurance (ESI). ESI often pay prices for medical products and services that are twice that of Medicare for the same service. And Medicare prices are themselves often higher than a lot of other countries.

Insurers have little ability, and little interest, in reining in prices. That's pretty simple economics - if you are a value-added-reseller of services (which insurers mostly are) and what you resell costs $10,000 with a 20% markup you make a lot more money than if you resell something that costs $1,000 with the same 20% markup. In insurer-speak, as long as they can raise premiums ahead of "medical trend" (i.e. increasing prices for medical services) they make more money.

Most employers have no idea how this works and even less ability to control costs. And since most insurance is indirect compensation (i.e. people only see the small fraction deducted from their paycheck, not the larger amount the employer pays, which is a form of employee compensation), employees have no idea how much they are paying.

One of the best things we could do would be to dismantle employer-sponsored insurance and give everyone Medicaid - i.e. a restrictive and basic plan that paid low prices for services and gave everyone access to essential medical care. What would likely happen is an outside market (these exist in many countries) that people with more money could use to get faster services, newer technologies, etc. There a lot of wealthy people in the US; this would lead to a robust post-insurance market but not the insane levels of subsidization we have now.

There would still be lot of lobbying by the medical industry, but since Medicaid is a state-federal partnership and some of the money comes out of state budgets, there would be more opportunities for cost control

What is the job of this assistant?
If it's to recognize simple tasks and put them in a task list, we're just about there and the stuff you describe is about improve the assistant's ability to communicate with you.
If it's to start doing the task, the assistant will rapidly run into problems of decision-making. Should it buy you ice cream at the store because you're feeling a little down and could use some ice cream, or should it buy you kale because you haven't really done as good a job eating your vegetables. Is that a setting in the AI? Should it do some kind of emotional inference from other data like facial expressions? Will that be a good experience or will you feel like you're fighting the algo (my bet is the latter).
If it does get ice cream, what should it do if Haagen Daz is out of stock? Should it use store brand instead?

These decisions frequently have an emotional component to them - by making them you feel differently than before you made them. That process-oriented, state-modifying nature of decision-making is going to really ramp up once AI stops writing blog posts nobody cares about and starts trying to actually make decisions for you.
Similar problems exist all of the space encompassed by the term "intelligence".

Current AI, for all its cleverness, is basically regression. As a number of AI experts have noted, the work on inference and reasoning basically got stalled when progress on the neural network approaches started to take off.

The problem is that this approach assumes that there is clear, unambiguous, objectively definable truth that can be used to define a training set for the AI. In reality, many if not most interesting problems, and certainly the hard ones, do not lend themselves to this at all. For example, imagine training an AI on the scientific literature of the past 100 years. Much of that literature will be considered wrong by present standards, and much of the rest will be small-scale and speculative. The truth isn't something that exists objectively, it's something that we construct out of a combination of verifiable facts, philosophical and epistemological frameworks, our own biases, our own emotions, and often randomness.

It is possible that a general AI could emulate all that, but there's a pretty decent chance that that would bind that AI to all the problems and biases that exist in human intelligence. And we know almost nothing about other intelligences, like what and how dolphins or elephants take hold of the world. We've mostly assumed away that concern by counting on historical dismissal of these beings' intelligences.

My guess is that AI will rapidly start to go in circles. It's pretty much already consumed much of human writing and still has no concept of truth whatsoever. This is likely to lead to a torrent of bullshit - basically spam in everything that will make it that much hard to engage in truth-seeking and truth-making.

It may get better some things that involve searching parameter spaces and combinatorics; that will doubtless be useful.

I just am not convinced that reality, knowledge, and epistemology actually lend themselves to the kind of AI that people are envisioning.

So there is some logic to having a container with no outbound network connection run the code. Excel does not have the same broad catalog of 3rd party libraries that Python does. So even though VBA has definitely had security issues, it is not a bad idea to limit Python's execution to a container.

That's where reasonable ends and the annoying part starts. I can't understand why Excel can't instantiate the same container locally, with the same restrictions on outbound traffic. At a guess, they didn't want to deal with managing the containers, and Anaconda is looking for ways to get license revenue.

The other problem is that when the data frame is instantiated in Python, it is encapsulated as an understand set of data. Allowing that data frame to be edited by the Excel UI opens up pretty complex scenarios of concurrency and confusion about data lineage.

There are definitely times when one wants to do something to a data frame, then pass it on for manual editing (often to another user), then resume working on it. On the whole that kind of thing is much easier to keep track of in systems like DBT or Dremio.

This is a great point. I am something of an EV enthusiast, at least insofar as I'm excited about the transition to electric. I drove the Tesla and enjoyed its zippy acceleration but wanted nothing to do with its tech bro bachelor pad interior. I ended up with a Pacifica PHEV minivan, which I like a lot. If it had 50 more miles of range (currently has 35 mi) I'd be ecstatic about it.

So I'm totally with you that the market has not yet filled in on styles and models that a lot of car buyers still want. The next two years are set to be pretty big ones, because all the legacy manufacturers are currently in the middle of transitioning to EVs. I suspect it will take a few more years after that before all the current body and interior styles become available as EVs.

As far as performance and value goes, I think in 5 years EVs will simply be better value than almost any ICE vehicle because the batteries will be at least as good as they are now, manufacturing costs will have come down, while the fueling and maintenance costs of EVs are 30-50% lower. It's conceivable that batteries will have gotten significantly better or cheaper, meaning that the value equation will be heavily skewed towards EVs.

The charging infrastructure will remain a pain point for years. The maze of charging standards and speeds, plus broken chargers, different vendors and accounts will make public charging a huge pain, while every single gas pump will still fit in every single gas tank. Those with a driveway or garage can basically stop thinking about fueling their car (a huge perk of EVs), while anyone who parks on the street or is in a multi-tenant property is likely to have a much harder time. Since this involves lots of third party property owners, I suspect it will take quite some time before everyone has a convenient place to plug in their vehicle at home.

I don't buy the "grid will collapse" FUD, but I do think in some places EVs will aggravate problems like brownouts in summer etc.One the flip side, if a decent number of EVs end up with vehicle to load, there'd be a moveable fleet of power in the event of storms, etc.

I'd keep your eyes on China - EVs have reached ~25% of new car sales there, too. I have to believe that this means there will be a much wider variety of styles made to meet the variety of consumer tastes.

There is a logic to trying to make regulations lighter on small entities, but it is not costless.
To frame this in engineering terms - this system design introduces a new set of dependencies , e.g.:
- Evaluation of some "optimal size" where regs grow stiffer
- Metrics for measuring a business' size relative to that standard
- Auditing that those metrics are reported honestly.

There are times when it's worth it - e.g. I've had to comply with HIPAA at a startup and many of the requirements seemed nonsensical for 10 people in a little office rather than a health system of 10k+ people, but carveouts add complexity quickly.

This is a great comment!

Much of the supposed debate about regulation vs deregulation seems to happen at the broad principles level, which makes it mostly noise. If you look a regulations as establishing a system of permitted/incentivized and non-permitted actions, these kinds of examples of perverse incentives abound and the need for better system design is glaringly obvious.

Nobody thinks it's possible to run a country without taxes but that doesn't make it obvious how to run an effective taxation system, much less wading into the thorny questions of fairness. The hard part is when you have to decide how much each citizen should pay and enforce it.

In the US we have several interlocking problems:
1- people are much more fond of debating abstract principles than the particulars. To me a good example of this is how the debate over taxation rarely ever raises taxation measures like VAT that are used successfully in many other countries.

2 - Another problem is that the US regulatory development process seems to involve very little a priori evaluation of how the system created by the rules will be used and abused. In cybersecurity terms the US tax code has an extraordinary amount of attack surface - at least in part because since the 70s we Americans have disguised a great deal of ordinary fiscal spending as tax expenditures. The low level of securing of this system would make any responsible security analysis blush. One upside is that this haphazard design was largely responsible for repelling attacks on our voting in the past few elections.

3- Still another is that we have a highly fragmented system for developing laws and regulations, thanks to the division of power between federal government, states, and local governments. That plus political stalemate (back to #1) means that fixing poor system design is also extremely hard and slow.

I wish we could somehow move our debate away from the pointless duck season wabbit season stuff and toward *how* to regulate effectively. Not holding my breath though.