100% and even if they might not "replace" a senior or staff SWE, they can make their job significantly easier which means that instead of requiring 5 Senior or Staff you will only need two.
LLM WILL change the job market dynamics in the coming years. Engineers have been vastly overpaid over the last 10 years. There is no reason to not see a reversal to the mean here. Getting a 500k offer from a FAANG because you studied leetcode for a couple weeks is not going to fly anymore.
“which means that instead of requiring 5 Senior or Staff you will only need two”
What company ever feels they have “enough” engineers? There’s always a massive backlog of work to do. So unless you are just maintaining some frozen legacy system, why would you cut your team size down rather than doubling or tripling your output for the same cost? Especially considering all your competitors have a similar productivity boost available. If your reaction is to cut the team rather than make your product better at a much faster rate (or build more products), you will likely be outcompeted by others willing to make that investment.
This i what is so tricky with Self Driving. People "feel" it is almost there because most of their rides are mostly ok. However to make a system truly driverless you need to master the long tails of difficult events and FSD is nowhere even close to do that.
Going from 99.9% to 99.99999% is what makes a system truly driverless and where most of the work is. Waymo is way way ahead of FSD for this.
I know essentially nothing about machine learning, nor about what approaches to ML that Tesla or Waymo are using. Is there a Metcalfe-type effect possible here? Where the better FSD becomes, the more people use it, finding more edge-cases, so the driving get better, so more people start to use it. With the end result that the learning starts to get better/faster with time, in a positive-feedback like mechanism?
Your comment/question seems to assume that somehow accurate classification of performance is possible, which is demonstrably false assumption.
First, if somehow system-internal watchdog was able to detect erroneous outputs for training, it would also be able to stop those outputs propagating to control system live, leading to zero errors. Such a watchdog requires self driving to be quantified analytically, leaving implementation entirely testable and therefore control system not passing the testsuite (i.e. exhibiting erroneous behavior, i.e. still having "edge cases") not deployable in public. Many words to say that that in practice with driving being somewhat loosely defined you need humans in the training loop.
Second, vehicle operators are not only unqualified to accurately monitor behavior of these autonomous systems, in part due to not being formally trained on safety systems in general and the system they are monitoring in particular, but on top of that incentivized to underreport erroneous output.
My prediction has been that we are going to see increased number of accidents with ADAS deployments, not less, before the number can start dwindling.
I worked in this industry. No, FSD is not almost there. Not even close. What matters is the long tails of events.
You might "feel" it is almost there because it gets it right 99.9% of the time but that is still way too many accidents and injuries in the long run. And the work to go from 99.9% to 99.9999% is 1000x more complicated.
We need to compare against the real human error rate (including drunk drivers, sleepy drivers etc). What is that error rate?
Also, fault tolerance error rates don’t work that way - difficulty increases exponentially as you increase fault. In other words, it’s much more difficult than three orders of magnitude to go from 3 9s to 6 9s - it’s easily 5-6 orders of magnitude.
When you're playing candy crush and FSD kills someone, whether or not it's drives better than the average driver is not something the judge and prosecutor are going to consider.
First, that’s because FSD is legally only L2 and thus legally you are required to pay attention. It has nothing to do with the engineering realities of the impact on improving safety.
When L5 becomes available then this becomes a different calculus. And it’s honestly questionable about how good the characterization actually is since Tesla’s L2 FSD seems to outperform Mercedes’s L3 and it’s more a matter of the liability the manufacturer is willing to take on vs objective measurements of quality.
yes, the long tail of difficult event is exponentially more difficult to handle. That's why I said above that people "feel" it is ready but it is nowhere to be even close to ready.
The average crash rate for human is one every 500k miles.
100 Miles literally don't matter. Even 1000. On average accidents happens every 500k miles.
Are you ready to have your tesla drive FSD with you sleeping in it for 500k?
You are letting your feeling dictate that FSD is ready. The math is more complicated.
Got it. differing expectations. Im not expecting 'unsupervised' out of the system, I dont think I would trust any system to transport me or loved ones un-piloted. Even when being driven in a taxi, I am still supervising despite not having much recourse beyond barking at the driver.
That said, based on my experience with FSD, I'd be tempted to take a nap if it let me, reinforcing my initial statement that it's very darn close.
If you drove with a very very bad driver that crashes every 50k miles (average is every 500k miles), you would have exactly the same feeling and you would also be tempted to take a nap.
> Brecht says his lab members didn’t initially believe Anchali was smart enough to deliberately sabotage Mary. But several things convinced them that the pachyderm knew exactly what she was doing. Kinking the hose is a complex behavior that requires a lot of twisting and pressure. Over the weeks in which the researchers were studying Mary’s showers, Anchali began to kink the hose more frequently and got better at it, suggesting she was practicing and learning. She also started to stand on the hose—something the caretakers had specifically trained the elephants not to do.
exactly this. Working with a couple ex-google and ex-facebook. Without failing every meeting we hear at least once "At Google/Facebook we used to ..." for something that is completely not applicable here.
I work at one of the big companies and some people do that even here. People just talk about their experience or other systems that they know. It always feels conservative to me, as often it's said to suggest emulating what some other group or company did to solve an often superficially similar problem, but with entirely different constraints. I think some people with certain kinds of thought patterns just pattern match and try to apply past experience a bit too broadly.
Ah, another one of those infamous 2018 blog posts on "why I quit Google".
And those blog posts absolutely always start by telling you that the engineers at Google are the smartest in the world. Oh boy are those people indoctrinated.
What sounded like the usual Google-internal-self-congratulatory-echo-chamber nonsense grated on me, too, but I kept reading, and was glad I did, because the article didn't go like the usual.
They were in 2008. It's largely just people's mental models changing slowly, as well as selection bias of people who still believe Google has world-class engineering being overrepresented among people who still work at Google.
2017-2022 I would've said the Ethereum foundation and various DeFi startups. If you look at what Ethereum accomplished with proof of stake, it's quite remarkable. The game theory alone for proof of stake is thorny as hell, and then they built a working secure implementation and upgraded a distributed network of unaffiliated organizations who all voluntarily run nodes, all without major hitches or anyone's funds being lost. That's huge, and more impressive than anything Google's done in the last 5 years.
Since 2022, I would probably say the various LLM owners: OpenAI, Anthropic, Character.AI, etc. The complexity here isn't in the basic transformer architecture though (which was invented by Google in 2017), but in the data infrastructure that hoovered up all the data on the web to train it on, as well as the fine-tuning on various use-cases.
Not surprisingly, good engineers follow the money: these have also been the lucrative industries in those time periods.
GDPR and CCPA etc made it easy to send a request for deletion that will most probably be a frontend gimmick. How much effort are they really going to put into going back in their backups and deleting all your entries? I'm pretty sure it must be the lowest roadmap priorities.
And it's amazing how financial liability has a way of getting things on a VP's feature radar that common sense doesn't.
The reason it was haphazardly handled prior was that there was no liability. Who cared? (legally speaking)
From working inside a T25 American retail company, I can say that we went top-to-bottom and rearchitected for traceability and hard deletes as a result of the CCPA.
I have a feeling that it's also quite a difficult problem past some scale of infrastructure.
If I ask Google to delete my data (EU citizen), I have trouble believing that they actually go through all of their cold storage backups where it was stored and make sure it's erased. At best I could believe that the process is designed in such a way that my soft-deleted data is unlikely to be recovered (intentionally or not) and maybe unlikely to be possible to link to my account.
What they should do (I have no idea what they do) is to encrypt every record belonging to a user with an individual key. Live records, backups, everything. If a user wishes to be deleted, that live key is simply obliterated, making any data the user owns unrecoverable.
Since the key is not used for end to end encryption, and backends still have access to the data (as long as the key lives), it has different requirements on how it needs to be protected. The biggest challenge is backing up the key itself, as losing it means losing access to all the user’s data by design. But backing up and obliterating a single key is much, much easier than doing so for a whole set of loosely associated data across many databases.
Practically speaking, it also makes using and querying that data and doing any kind of analytics much, much more expensive. It is done that way in some cases, but in the absence of a technical requirement to do so, there are cheaper approaches.
Those are solvable problems. I could also argue how address space separation and more generally MMU protections make things so, so much more complex (they do!), yet we don’t question that one very much.
There is no end to end encryption involved here, so you don’t need to resort to such voodoo as homomorphic encryption.
> the problem only partially smaller, since you still need to sync and back up the keys.
I mentioned that: It makes the problem much smaller, as you only have one single, small piece of data to backup and and erase, instead of an ever-changing many-faceted blob of distributed data.
> Also, is an encrypted piece of data with a lost key truly deleted? What if the encryption gets cracked?
Oh boy. If simple symmetric encryption gets “cracked”, then you have much larger problems.
> I would say it is more deleted than toggling a `deleted` flag in the db and less deleted than burning the tapes in fire.
For all practical purposes symmetrically encrypted data that lost its keys is considered “random” data. If you “erase” data on a device before you sell it, most often it will just throw away the key to the disk contents nowadays.
They already do this (the encryption-at-rest part). Deleting the data is still a hard requirement. Also, the keys are never seen outside of the centralized encryption service. Deletion is still a must.
Before you make a deletion request, make a subject data request and see what they have on you; then request deletion; then make a subject data request again.
Google-scale companies have very capable people employed, both on the technical and legal side, who do nothing else than look for these kinds of oversights, and are empowered to make sure they get fixed.
Large companies fail in spectacular ways all the time. Google is super successful because they tapped into the biggest cash cow of all times. Not because the employees are somehow very capable and above any oversight.
I can't speak for any other companies, but you don't need to speculate. You can search the internet and find several articles outlining that the correct strategy for businesses here is to delete the data from production systems, and then maintain a record of references to those deleted records such that a restored backup can ensure that deleted records are not put back into production.
There is generally an expectation that data may be retained in backups for a specified retention period, but will not be used or restored. Beyond that, it is up to the regulator to determine if this is meets the standard, but it's worth noting that there are notions baked into the text and the interpretations of the text of GDPR that account for reasonable costs and efforts.
Auditors can and do test and monitor for this, both using audit processes and demanding evidence, and by performing manual testing and experimentation.
Fines for non-compliance with GDPR regarding data of European citizens can amount to 4% of annual revenue:
83(5) GDPR, the fine framework can be up to 20 million euros, or in the case of an undertaking, up to 4 % of their total global turnover of the preceding fiscal year, whichever is higher.
Exactly this. Especially for a currently failing company that got an incentive to NOT delete your data (because that's the only value they still have).
Technically true but this feature never had much update and I believe hdmi ARC and this are mutually exclusive/use the same wiring path?
Anyway I think the real threats will be:
1) Aggressive wifi search connecting itself, including deals with ISP routers to allow them to bypass you or even other devices.
2) Time-bombs causing the TV to become non-functional or degraded if you don't connect it to the internet, after the warranty or return window has expired
3) In-built 5G modem connectivity (everyone says this is to expensive but manufacturers could cut bulk deals and could limit the bandwidth usage, even just sneaking in firmware updates has a lot of abuse potential)
Almost nothing supports HDMI Ethernet, it is the use of extra signal path on the cable to provide an ethernet link between two devices. Both devices have to support the extra signals and one of them has to be able to route, so it's basically someone's bright idea that barely has any support.
2) I've designed a system like this for a TV rental company, although legally a general retail TV manufacturer wouldn't want to do this because it causes reputational damage.
3) There's material cost for the modem and then there's a subscription cost for the connectivity. Either of which would reduce their profitability.
I spent a decade working in consumer electronics, working with all the major brands you know well, many of the white companies who make the components and the ODMs who make the boxes that will get a brand stamped on them by whoever is buying it.
Ultimately the TV business is barely profitable, most big brands sell TVs as something of a loss leader so that they can sustain their brand name. You spend each night with a Samsung, LG or Sony remote in hand looking at their product? Then they're winning in their eyes. Also because of the relatively high value of the TV it sustains their overall turnover without actually contributing to profitability. When a manufacturer launches a new TV they get about 8 months to make a profit on it, after that they're probably losing money because of downward pressure by retailers to drop the price. That's driven by consumer demand for cheaper rather than better products by the way, consumers have some responsibility for the state of the market.
The Smart apps systems cost the TV manufacturer, they have to supply the servers and infrastructure. They may make a small commission if a customer signs up to a streaming service on their device, but otherwise your general use of Smart technology costs them money every day.
Ultimately, most TV manufacturers have zero interest in spying on you. LG's biggest blunders can all be traced back to a lack of care and due diligence in their handling of data. Most of the time the 'mass data collection' is just accidental, someone in the development team thought it would be a good idea to collect data and some researcher is horrified by how much data gets sent back. Sometimes, someone gets the idea that viewing data could be used to put ads on the product, but ultimately they're not interested in what you watch, they're interested in grouping you into an advertising bucket so they can suggest you watch another movie with a Hemsworth in it.
I'm not saying that there shouldn't be oversight, and that these companies don't do stupid things for money, but ultimately there's never malice, or a desire to spy. Most of the overreach is incidental to the overall goal.
If someone doesn't want to use smart TV tech, then I'd advise them to not connect the TV to the network. There are set-top boxes out there that can do the job easily enough, and some of them might not even spy on you. One thing to remember is that many Android boxes you buy online, especially the "IPTV" ones, are riddled with malware. So don't think that by disabling Samsung and going to Kodi, you're making yourself safer.
LLM WILL change the job market dynamics in the coming years. Engineers have been vastly overpaid over the last 10 years. There is no reason to not see a reversal to the mean here. Getting a 500k offer from a FAANG because you studied leetcode for a couple weeks is not going to fly anymore.