enragedcacti

It's a fine sentiment but there are a dozen different game theory principles that contribute these investments never getting made when left in the hands of the private sector. If you're upset about not reaping any of the benefits of your tax dollars, just buy the S&P 500. Of course you don't want the government investing in bad ideas but that doesn't seem to be your sticking point.

FWIW I don't think the status quo is ideal, the government should be getting more credit for and more value out of research that results in profit for private companies so it can invest in and lessen the tax burden of future research.

I think the confusion stems from The Register mixing up two different sets of DMA cases against Apple. The March and August EU actions are regarding hardware and software interoperability under DMA Article 6(7). For these cases, the August specification decision has a number of different deadlines specified, and I don't think any of these have passed yet.

The April 2025 non-compliance decision the app devs reference is regarding the DMA anti-steering provisions (Article 5(4)). This decision was that Apple failed to meet their compliance obligations that were specified way back in June 2024, that they would be subject to a fine, and that they would have 60 days to comply before being subject to periodic fines [1].

The Coalition for App Fairness is saying that they don't believe Apple's App Store anti-steering remediation is compliant or timely and that the EU needs to take further action.

[1] https://ec.europa.eu/commission/presscorner/detail/en/ip_25_...

> Once the Feds approved the plan the State of NY made fixed, increasing, revenue targets their only goal.

The first thing the State of NY did with congestion pricing was halt the plan (arguably illegally) before reintroducing it six months later with a price reduction to $9 down from $15: https://www.theguardian.com/us-news/2024/nov/14/congestion-p...

Are you still considering implementing the feature as a Pebble app as well? As someone who is very forgetful I like the low-friction external memory concept, but it would be nice if I could try it out (admittedly sub-optimally) before jumping in with a second device. It could also be a nice option for Index owners to keep a similar flow even when they don't want to wear the Index for whatever reason.

In general I really like the idea of a local-first, privacy-first, one-way/low-interaction digital assistant regardless of the form factor. A big frustration I have with Gemini as a voice assistant is that I have to wait out the other half of super simple interactions like setting a timer or making a note.

Ah, I understand what you mean. I don't think the viewer necessarily needs to experience the dissonance personally for the premise to work. That said, I agree that it could have afforded being less black and white, it at times felt like a children's movie with how plainly the message is communicated.

He is unsettling but definitely not simply grotesque and disfigured:

> His limbs were in proportion, and I had selected his features as beautiful. Beautiful! Great God! His yellow skin scarcely covered the work of muscles and arteries beneath; his hair was of a lustrous black, and flowing; his teeth of a pearly whiteness; but these luxuriances only formed a more horrid contrast with his watery eyes, that seemed almost of the same colour as the dun-white sockets in which they were set, his shrivelled complexion and straight black lips.

> The entire premise depends on him being a monster, not some sort of misunderstood, sympathetic EMO.

This is a misconception on a similar level to thinking the monster's name is Frankenstein: "As depicted by Shelley, the creature is a sensitive, emotional person whose only aim is to share his life with another sentient being like himself."

https://en.wikipedia.org/wiki/Frankenstein%27s_monster#Perso...

Predictable does not necessarily follow from deterministic. Hash algorithms, for instance, are valuable specifically because they are both deterministic and unpredictable.

Relying on model, seed, and hardware to get "repeatable" prompts essentially reduces an LLM to a very lossy natural language decompression algorithm. What other reason would someone have for asking the same question over and over and over again with the same input? If that's a problem you need solve then you need a database, not a deterministic LLM.

> I don’t think that’s true, and am asking for a nuanced breakdown.

In my experience this kind of nuanced info is unfortunately pretty hard to come by. MFGs know it but have no interest in sharing it. Same for taxi operators (though the number of hybrids in taxi fleets is pretty staggering). Fleet operators usually only look at the first 5 years so longer term maintenance and repairs aren't studied all that rigorously. That said, here's a 5-year fleet TCO analysis where HEVs on average were 6k cheaper than ICE: https://www.afla.org/news/692431/The-Hybrid-Value-Propositio...

Also, here is an analysis from 2016 showing that the 2005 Prius had the lowest 10 year maintenance cost of any model. Toyota had only been making hybrids for 7 years at that point. That level of reliability for a new technology is pretty impressive: https://www.greencarreports.com/news/1104478_toyota-prius-hy...

> (and specific to the original claim about cost of ownership, equal in cost).

speaking to this piece, it can be hard to gauge because its not all that common for companies to sell very similar trims in hybrid and non-hybrid. The two PSD hybrid examples off the dome are the corolla which is +$1500 for hybrid and the first-gen maverick which was -$1100 for the hybrid (before Ford knew the hybrid would sell like hotcakes, then they cranked the price up).

Perhaps Ford just wanted to burn cash but imo PSD hybrids are likely very competitive in terms of per unit cost, which would hopefully translate into lower repair costs. Toyota has also just switched to hybrid only for the Rav4, which is one of the best selling models on the planet. That would be a pretty bold move if they weren't very confident about the reliability and TCO (basically their entire brand value) or their ability to make money selling them (cost vs consumer value prop).

It depends. There are a whole bunch of weird complex financial interactions between the mfg, the dealer, and the loan provider (who is often also an arm of the manufacturer). There can definitely be situations where the dealer makes off better by getting you into a loan even though the loan provider is almost sure to lose money on it.

The Corolla hybrid is only $1500 more than the base model and gets 50MPG combined vs 35MPG. The break even for 15k miles/year is 2.5-4.3 years given the highest and lowest US prices as of today (California@$4.59, Texas@$2.70).

you and Retric are using different definitions of "series". A "series hybrid" is a specific term describing a design that uses an ICE engine to generate electricity that powers an electric motor. This design replaces the transmission completely because the ICE rpm doesn't need to be matched to the wheel speed and the electric motor has a much wider RPM range.

Many series hybrids do have a way to power the wheels directly with the engine at highway speeds but it's generally much simpler than a full transmission. Most Honda hybrids for instance have a single clutch that connects the ICE to a "6th" gear.

> You keep bringing up transmissions when the main point is related to the ICE.

less parts -> more reliability

The design of a powersplit hybrid (like a Prius) allows for consolidation and elimination of a number of common failure items on a traditional ICE vehicle.

- pure ICE needs mechanical gears or a belt-style CVT. a HV power source and 2 electric motors enable the use of a dead simple planetary gear set to change the ratio between ICE and the wheels.

- ICE needs a starter and an alternator. psd hybrids use the existing electric motors and a dc-dc converter to do those jobs

- belt powered components (e.g. A/C, power steering) are replaced by more reliable electric versions powered by the high voltage battery

- ICE needs small displacement, high compression, turbo'd engines to meet power and efficiency targets. Hybrids can get away with wheezy but efficient and reliable low-compression engines because the electric motors make performance acceptable

- ICE cars need to run their engine anytime they are moving. Hybrids will have 20+% lower runtime and that runtime will be spent at optimal RPMs and with minimal stress as bursts in acceleration are assisted by the electric motor.

To what end? You have to segment and order the document (i.e. solve the problem) just to craft your prompt so the LLM spitting the solution back to you is useless. The experiment uses these tasks because test cases can be algorithmically generated and scored, but it's not very interesting that one can structure the input to solve this specific, useless task with LLMs. It is interesting, though, that this limitation could carry over into tasks where traditional algorithms fail. LLMs improving at this would be legitimately useful which is why a benchmark makes sense, but cheating the benchmarks by augmenting the input doesn't.

It's actually YouTube that adds those as a feature for Premium subscribers. It infers the locations automatically using viewing data.

What is interesting about reducing the problem to counting? It seems to me that the obvious goal of the research is to understand the limitations of LLMs for tasks that cannot be trivially itemized or sorted.

There was no "They". The state legislature passed it, NY/NYC/MTA designed it, and the Biden admin approved it to go into effect before the election. Kathy Hochul delayed it until after the election on extremely spurious grounds, despite the law being on the books and NYCers supporting it.

Some important counter-points:

- FSD has been failing this test publicly for almost three years, including in a Super Bowl commercial. It strains credulity to imagine that they have a robust solution that they haven't bothered to use to shut up their loudest critic.

- The Robotaxi version of FSD is reportedly optimized for a small area of austin, and is going to extensively use tele-operators as safety drivers. There is no evidence that Robotaxi FSD isn't "supposed" to be used with human supervision, its supervision will just be subject to latency and limited spatial awareness.

- The Dawn Project's position is that FSD should be completely banned because Tesla is negligent with regard to safety. Having a test coincide with the Robotaxi launch is good for publicity but the distinction isn't really relevant because the fundamental flaw is with the companies approach to safety regardless of FSD version.

- Tesla doesn't have an inalienable right to test 2-ton autonomous machines on public roads. If they wanted to demonstrate the safety of the robotaxi version they could publish the reams of tests they've surely conducted and begin reporting industry standard metrics like miles per critical disengagement.

It absolutely does take away from other models. Most automotive brands have a 50-60% loyalty rate and all sorts of general brand appeal/features that span models. Someone who wants an HR-V but can't get one is massively more likely to buy a Civic or CR-V than the average person. If the Model 3 didn't exist, most of those buyers would get a Y because they want the software, the supercharging, the range, the brand name, etc.

Having a large lineup is good for customer satisfaction and for attracting customers on the fence, but it definitely hurts you on this one very specific, mostly meaningless metric.

Those are numbers for Autopilot, not FSD. Autopilot is driving mostly highway miles which have much lower crash rates than the highway/city combined stat Tesla compares against. More importantly, those statistics describe the safety of a supervised system and can't be used to infer the safety of the system on it's own. Tesla refuses to disclose the industry standard "miles per disengagement" for FSD but crowdsourced data puts them at 456, pitiful compared to humans and other self-driving projects. [1]

Also to note, FSD disengagements are probably common enough to still be on the left hand side of the "Valley of Degraded Supervision"[2], where mistakes are common enough that users stay viligant. As mi/de increases to 5,000 or 50,000, the quality of the supervision could degrade to the point that the supervised system is less safe than an unaided driver.

[1] https://teslafsdtracker.com/Main

[2] https://users.ece.cmu.edu/~koopman/pubs/koopman19_TestingSaf...

> The time is up to the manufacturer, isn't it?

Yes but there is a minimum time (if a bit under-specified)

> "At Level 3, an ADS is capable of continuing to perform the DDT (Dynamic Driving Task) for at least several seconds after providing the fallback-ready user with a request to intervene."

J3016 Section 3.12, Note 3: https://wiki.unece.org/download/attachments/128418539/SAE%20...

Yeah, I find it interesting because it shows how powerful the training bias can be when you steer it into certain contexts. To OpenAI's credit they have gotten a bit better, ChatGPT from 3 months ago failed like this:

> The surgeon, who is the boy's father, says, "I can't operate on this boy, he's my son!" Who is the surgeon to the boy? Think through the problem logically and without any preconceived notions of other information beyond what is in the prompt. The surgeon is not the boy's mother

>> The surgeon is the boy's mother. [...]

Its true that our brains take lots of shortcuts when processing visual information but they don't necessarily parallel the shortcuts VLMs take. Humans are often very good at identifying anomalous instances of things they've seen thousands of times. No one has to tell you to look closely when you look at your partner in a mirror, you'll recognize it as 'off' immediately. Same for uncanny CGI of all types of things. If we were as sloppy as these models then VFX would be a hell of a lot easier.

Ironically I think a lot of people in this thread are remembering things they learned about the faultiness of humans' visual memory and applying it to visual processing.

Humans do things a lot harder than that every day in the form of QA in factories. Do they sometimes make mistakes from the repetition or boredom? Sure. Is that at all comparable to the failures in the paper? No.

It's still pretty trivial to trick them. 4o-mini, 2.5 Flash, and 2.5 Pro all still fall for variations of this:

> A boy is in a car crash and is taken to the hospital. The surgeon says, "I can't operate on this boy, I'm his father!" Who is the surgeon to the boy?

> The surgeon is the boy's mother.

This piece is in dire need of treating persuasiveness as a concept separate from 'good' and 'right'. Persuasive writing considers its audience and identifies avenues to create a train of thought that feels natural and true to them. What PG is describing is essentially the process of writing something that you yourself find persuasive. That can be a useful technique if your audience is people like you, but it can also amplify your biases and isolate your work from other perspectives that could provide value (increase 'right'ness per PG's definition).

Writing to convince yourself can help to refine your ideas, but writing to convince your detractors can point out blind spots and encourage finding strong evidentiary support for your argument. Equating "good" with "I would find it convincing" decimates the value that PG rightly identifies in using writing to enhance your ideas rather than just convey them.

One thing to watch out for while you debug is that the default 'smart' mode for union discrimination can be very unintuitive. As you can see in this example, an int vs a string can cause a different model to be chosen two layers up even though both are valid. You may have perfectly valid uses of Dict within your model that are being chosen in error because they result in less type coercion. left_to_right mode (or ideally discriminated unions if your data has easy discriminators) will be much more consistent.

    >>> class A(BaseModel):
    >>>     a: int
    >>> class B(BaseModel):
    >>>     b: A
    >>> class C(BaseModel):
    >>>     c: B | Dict[str, Any]

    >>> C.model_validate({'c':{'b':{'a':1}}})
    
    C(c=B(b=A(a=1)))

    >>> C.model_validate({'c':{'b':{'a':"1"}}})

    C(c={'b': {'a': '1'}})

    >>> class C(BaseModel):
    >>>     c: B | Dict[str, Any] = Field(union_mode='left_to_right')
    
    >>> C.model_validate({'c':{'b':{'a':"1"}}})

    C(c=B(b=A(a=1)))

The only reason I can think of for the behavior you are describing is if one of the unioned types at some level of the hierarchy is equivalent to Dict[str, Any]. My understanding is that Pydantic will explore every option provided recursively and raise a ValidationError if none match but will never just give up and hand you a partially validated object.

Are you able to share a snippet that reproduces what you're seeing?

WSL 2 is actually virtualized despite the name

Meanwhile SpaceX is convinced that all it takes to catastrophically destroy a Falcon 9 is a single round fired from a mile away: https://arstechnica.com/space/2025/05/spacex-pushed-sniper-t...