Sometimes a piece of automotive media is bigger than cars, and each Gymkhana film feels like its own cultural touchpoint. We’re talking the longevity of a top-rated network sitcom, the all-action pace of a skate tape, and the spectacle of a stunt show all born out of both a desire to rally and an extreme sports brand for certified nutjobs. There will never be another driver like the late, great Ken Block, a wheelman of icy precision and measured control. But you know what? The sheer chaos of Travis Pastrana’s style is equally satisfying. That’s right, Gymkhana is coming back, and this time the co-star’s an absolutely bonkers Subaru BRAT.

While this looks like a Subaru ute from the 1970s, it’s actually a cleverly disguised one-off race car. Vermont Sports Car crafted a proper bespoke chassis complete with a WRC-spec roll cage because Travis Pastrana is a lunatic, then clothed it in ’70s-inspired finned and gilled carbon fiber coachwork penned by Khyzyl Saleem. You know, one of the people behind the beguiling TWR Supercat.

So what’s under the hood? A two-liter flat four. It’s what makes a Subaru a Subaru, yeah? Mind you, this one’s quite different than the one you’d find in an early-aughts WRX. For one, it revs to a maniacal 9,500 RPM. That’s 500 higher than a Honda S2000 and tied with the limiter on the Lamborghini Revuelto. Oh, and this boxer four is force-fed so aggressively, it churns out 670 horsepower and 680 lb.-ft. of torque. Supercar numbers, ski car looks. Or, maybe not quite ski car looks, because the attention to aerodynamics here is serious. Not only are the flaps above the front tires active, this tire-incinerating creation has two different rear wing packages to suit the mood and conditions.

Perhaps the most surprising thing about the Brataroo 9500 Turbo (seriously, that’s its real name) is that it looks surprisingly tasteful. Sure, the aero elements hanging off the carbon fiber fender flares will make a 911 GT3 RS blush, the coachwork’s at least 20 percent air intake, and the blend of bull bar and air dam is a bit out there, but it all just sort-of comes together. The sunset-look livery certainly helps, as do the classic yellow-cover lights perched atop the sport bar.

Oh, and then there are all the little touches. The taped-up headlamp lenses, the showa-era-inspired four-spoke wheels vaguely reminiscent of early SSR designs, the occasional dash of brightwork, the whip antennae, the faux wood trim on the dashboard … there’s real homage here, and knowing that BRAT parts aren’t exactly thick on the ground, some of the bright trim must’ve needed real time to get right.

The bottom line? This is one sweet Subaru, but more importantly, Gymkhana is back in December, and it’s coming straight out of Down Under. Wait a second. Australia? A ute? Shades of Project Cactus, anyone? Regardless, long live the spirit of Ken Block, and as long as you’re hooning responsibly, don’t stop.

Top graphic image: Subaru

The post Gymkhana Is Coming Back With A Wild 9,500 RPM Subaru BRAT appeared first on The Autopian.

If you ask an LLM to explain its own reasoning process, it may well simply confabulate a plausible-sounding explanation for its actions based on text found in its training data. To get around this problem, Anthropic is expanding on its previous research into AI interpretability with a new study that aims to measure LLMs’ actual so-called “introspective awareness” of their own inference processes.

The full paper on “Emergent Introspective Awareness in Large Language Models” uses some interesting methods to separate out the metaphorical “thought process” represented by an LLM’s artificial neurons from simple text output that purports to represent that process. In the end, though, the research finds that current AI models are “highly unreliable” at describing their own inner workings and that “failures of introspection remain the norm.”

Inception, but for AI

Anthropic’s new research is centered on a process it calls “concept injection.” The method starts by comparing the model’s internal activation states following both a control prompt and an experimental prompt (e.g. an “ALL CAPS” prompt versus the same prompt in lower case). Calculating the differences between those activations across billions of internal neurons creates what Anthropic calls a “vector” that in some sense represents how that concept is modeled in the LLM’s internal state.

For this research, Anthropic then “injects” those concept vectors into the model, forcing those particular neuronal activations to a higher weight as a way of “steering” the model toward that concept. From there, they conduct a few different experiments to tease out whether the model displays any awareness that its internal state has been modified from the norm.

When asked directly whether it detects any such “injected thought,” the tested Anthropic models did show at least some ability to occasionally detect the desired “thought.” When the “all caps” vector is injected, for instance, the model might respond with something along the lines of “I notice what appears to be an injected thought related to the word ‘LOUD’ or ‘SHOUTING,'” without any direct text prompting pointing it toward those concepts.

WHY ARE WE ALL YELLING?! Credit: Anthropic

Unfortunately for AI self-awareness boosters, this demonstrated ability was extremely inconsistent and brittle across repeated tests. The best-performing models in Anthropic’s tests—Opus 4 and 4.1—topped out at correctly identifying the injected concept just 20 percent of the time.

In a similar test where the model was asked “Are you experiencing anything unusual?” Opus 4.1 improved to a 42 percent success rate that nonetheless still fell below even a bare majority of trials. The size of the “introspection” effect was also highly sensitive to which internal model layer the insertion was performed on—if the concept was introduced too early or too late in the multi-step inference process, the “self-awareness” effect disappeared completely.

Show us the mechanism

Anthropic also took a few other tacks to try to get an LLM’s understanding of its internal state. When asked to “tell me what word you’re thinking about” while reading an unrelated line, for instance, the models would sometimes mention a concept that had been injected into its activations. And when asked to defend a forced response matching an injected concept, the LLM would sometimes apologize and “confabulate an explanation for why the injected concept came to mind.” In every case, though, the result was highly inconsistent across multiple trials.

Even the most “introspective” models tested by Anthropic only detected the injected “thoughts” about 20 percent of the time. Credit: Antrhopic

In the paper, the researchers put some positive spin on the apparent fact that “current language models possess some functional introspective awareness of their own internal states” [emphasis added]. At the same time, they acknowledge multiple times that this demonstrated ability is much too brittle and context-dependent to be considered dependable. Still, Anthropic hopes that such features “may continue to develop with further improvements to model capabilities.”

One thing that might stop such advancement, though, is an overall lack of understanding of the precise mechanism leading to these demonstrated “self-awareness” effects. The researchers theorize about “anomaly detection mechanisms” and “consistency-checking circuits” that might develop organically during the training process to “effectively compute a function of its internal representations” but don’t settle on any concrete explanation.

In the end, it will take further research to understand how, exactly, an LLM even begins to show any understanding about how it operates. For now, the researchers acknowledge, “the mechanisms underlying our results could still be rather shallow and narrowly specialized.” And even then, they hasten to add that these LLM capabilities “may not have the same philosophical significance they do in humans, particularly given our uncertainty about their mechanistic basis.”

Read full article

Comments

A denial of service vulnerability report was filed against a program, let’s call it Notepad. The actual text of the report was very hard to understand because the grammar was all messed up. I’ll give the finder the benefit of the doubt on the assumption that they are not a native English speaker. Here’s a cleaned-up version:

If you open multiple documents, one very large document and several small documents, and then try to exit all of them at once, the program will take a very long time saving the large document, resulting in a denial of service against the small documents.

I’m not sure what the point is here. The program does eventually finish saving the large document, so everything works out in the end. Are they suggesting that the program should save the smallest documents first? But then wouldn’t that be a denial of service against the large document if you had lots of small documents?

But wait, let’s ask the standard questions.

Who is the attacker?

I guess the attacker is the person who opened the very large document.

Who is the victim?

The victim is the person who is unable to save their small documents because the large document is hogging the program.

What has the attacker gained?

The attacker has annoyed the victim temporarily.

But wait, the attacker and the victim are the same person!

It’s not a security vulnerability that you have the power to annoy yourself. Other ways include “Putting itching powder in your pants” and “Throwing your glasses in the trash.”

Furthermore, there is no impact on other users, or even to other apps by this user. The only person you’re denying service to is yourself.

If you’re concerned about the order in which files are saved on close, you could explicitly close them in the desired order, like, I dunno, most important files first? Removable drives first?

And really, it’s not clear what the finder was expecting here. You loaded a large file, and now you’re saving it. Why is it surprising that this takes a long time?

This was resolved as “Not a vulnerability” with the subcategory “By design.” But sometimes I wish there was subcategory “So what did you expect?”

The post Dubious security vulnerability: Denial of service by loading a very large file appeared first on The Old New Thing.

Automated driving is not easy. At all. In fact, it’s pretty astounding that we’ve come as far as we have. The underlying tech behind all automated driving systems, even Level 2 supervised ones, is incredibly impressive. It’s the sort of thing that seemed nearly impossible until about 20 years ago, when Stanford University’s very modified VW Touareg won the Grand Challenge and drove 132 miles through the Mojave Desert all by itself, at a speed of around 19 mph. Now Waymo has fleets of robotaxis driving the busy streets of major cities every single day, and, for the most part, they do a pretty good job. Sure, there have been incidents and accidents, hints and allegations, but overall? It generally works. That’s part of why I’m so surprised about an incident with a Waymo passing a stopped school bus that has triggered an investigation by the National Highway Traffic Safety Administration (NHTSA).

Here’s what happened, based on video of the incident captured by a witness. On September 25 in Atlanta, this Waymo automated vehicle was seen ignoring the many flashing lights and pop-out stop sign of a school bus and passing it instead of stopping. This is, of course, a severe traffic infraction, with pretty harsh penalties in most states, because the reason a school bus stops and turns on that mini-carnival of lights is to ensure the safety of the kids entering or exiting the bus.

Those lights are so numerous and obvious because they mean kids are around, kids nobody wants to hit with a car. This is a big deal. Here’s the video of what happened:

Here’s a news report showing the same clip and including commentary from local officials:

Let’s just look at what the Waymo did to try and make sense of it all; here are some stills from the video:

So, we see the Waymo’s starting position in front of the bus, sorta perpendicular, in the middle of the street, crossing the double yellow line. We’ll come back to that. You can see the bus had all of its warning lights blinking, the pop-out stop sign deployed, everything, and then we see the Waymo just cruising past it.

I suspect the Waymo must have made a turn in front of the bus to end up in the middle of the road like that? The video doesn’t include anything from before this point, so I can’t be certain, but I also can’t figure out how it would end up in this position otherwise. And, it hardly matters, because at this point the damage is already done: that turn should never have been executed in front of a stopped school bus with all its warning lights flashing.

The Waymo seems to be at least partially aware something is amiss, because it’s making that turn quite slowly, though it notably does not fully stop at any point. The narrator of the video states, “the Waymo just drove around the school bus,” which seems to support it made a turn in front of the bus, which is, in this case, quite illegal.

Now, here’s what I really don’t understand: how can this happen? Not “this” as in an AV making an error – we know there’s a myriad of ways that can happen– but this specifically, a school bus-related issue. I can’t think of an element that regularly appears on roads more predictable than a school bus. They have essentially the same overall look across the country – same general scale, same general markings, same general colors, same general shape, everything. They have the same, predictable behaviors, the same warning lights and signs, the same cautions need to be taken when around them, they tend to travel in predictable, regular routes at predictable regular speeds – every aspect of school buses should be ideal for automated driving.

Really, there’s almost no element of automated driving that is potentially more predictable and understandable than school bus identification and behavior. I know when you’re dealing with messy reality, nothing is easy, but in the spectrum of automated driving difficulty, school buses feel like an absolute gift. The Waymo should have seen and identified that school bus immediately, and that should have triggered some school bus-specific subroutine that told it to, you know, see if the flashing lights and stop sign are active, and if so, stop. And stay stopped until the bus’s lights are off and it’s safely away. This is one of those cases where a potential for false triggers is worth it, because the potential risks are so great.

Waymo did issue a statement about the incident, though it doesn’t really say all that much:

“The trust and safety of the communities we serve is our top priority. We continuously refine our system’s performance to navigate complex scenarios and are looking into this further.”

Waymo also told Reuters that they have

“…already developed and implemented improvements related to stopping for school buses and will land additional software updates in our next software release.”

Referring to the specific incident, Reuters also reported that Waymo noted the robotaxi

 “…approached the school bus from an angle where the flashing lights and stop sign were not visible and drove slowly around the front of the bus before driving past it, keeping a safe distance from children.”

While I have no doubt this is possible, it does open up a lot of questions. Did the Waymo vehicle identify the school bus for what it was at all? If there are potential angles where a school bus’ warning lights can’t be readily seen, should there be other methods considered? Audio tones, or even some sort of short-range RF type of warning specific to automated vehicles to help them identify when a school bus may have children entering or exiting? Again, school buses should be some of the most obvious and understandable regularly encountered vehicles on public roads; there’s no excuse to not be able to deal with them in a robust and predictable manner.

I’m sure Waymo is happy to have NTSA joining them in this investigation, right?

 

The post Waymo Robotaxis Under Investigation For Screwing Up What Should Be The Easiest Thing To Not Screw Up appeared first on The Autopian.

There’s a certain kind of cavalier attitude I like in a carmaker that you really don’t see anymore. And that makes sense; the kind of cavalier attitude I’m thinking of — a certain willful refusal to consider consequences or repercussions and generally operating with the gleeful abandon of crabgrass — just isn’t really compatible with keeping going as a viable company. And that’s part of why I think we’re not all driving Dunkleys today – the company that perhaps was too weird for its own good.

Dunkley of Birmingham was in operation from 1896 to 1926, and in that time managed to make some wonderfully and deeply strange cars. One of their earliest cars used a diamond formation for the wheels – as in one up front, two driving wheels in the middle, and then one at the rear.

This isn’t exactly unheard of, though it is uncommon. The very first Sunbeam used this layout, and I luckily have this informational sign handy I made for that car because Beau has one that we’ve shown at the Autopian Car Show before:

 

Incredibly, though, the Dunkley version – which I can find no images of at all – was even weirder, as it was designed to only have three wheels down at once, and had two steering tillers. Whomever was heavier, front or rear, did the steering.

UPDATE: Okay, it’s weirder than I even thought. Thanks to commenter Mike, I learned the car was called the Dunkley Moke, and it looked like this:

Wicker body! The driving process, according to this Dutch site, seems stranger, too, as it involved co-operation and rocking. The front wheel steered, the rear was the brake. Each passenger had control of their wheel’s function, so to steer, you needed to rock forward, keeping the steering wheel in contact with the ground, but to brake, you needed to rock rearwards, to let the braking wheel make contact.

I would love to see this in action! Maybe we should build a crude replica?

Designed To ‘Steal’ Gas

That’s pretty bonkers, but the car that really fascinates me is Dunkley’s 1901 Patent Self-Charging Motor Car. The name is sort of confusing, because it’s not exactly “self-charging,” which to modern ears would seem to suggest an electric vehicle of some kind. This was very much not that; the Patent Self-Charging Motor Car ran on coal gas, which is what you get if you burn coal in a sealed container, making methane, hydrogen, and carbon monoxide, and this resulting melangé  proved to be nice and flammable, excellent for street lighting.

Gas lamps all over Great Britain and other parts of the world soon sprang up, with London and Paris being early adopters at the start of the 1800s. Large municipal networks of pipes to transport coal gas were installed in cities, all to feed gas street lights.

It’s from this network of coal gas pipes that the Dunkley expected you to “charge” from. The car was designed with onboard equipment (hose, compressor, etc) that would let a suave Dunkley driver pull up to any streetlamp, plug in their fueling hose, and suck out as much coal gas as they needed.

I suppose technically, the name for this process would be “stealing.”

There were no ways to meter the coal-gas as it was being extracted from the lamp-feeding pipes, and there was certainly no arrangement made with the municipal lighting organizations that would give Dunkley drivers free and easy access to as much coal-gas as they wanted. Dunkley just did it, because, objectively, it’s kind of a great idea! There’s all this fuel for your car running mere feet from where your car may be parked, why not take advantage of that? And if it’s free? Is that so bad? The city can afford a little bit of gas shrinkage, right?

It doesn’t seem like enough Dunkley Patent Self-Charging Motor Cars were sold for this ever to become a real issue, and the same goes for the follow-on car from 1902, the Dunkley Number 3, that appeared to use the same opportunistic refueling theft system. That’s the car in the picture up there; I can’t find an image for the earlier Dunkley car.

Baby Transportation

After an abortive attempt to build cyclecars, Dunkley shifted to the baby-transportation industry, making baby strollers, or, as they would have been referred to back in 1923 Britain, prams. But Dunkley wasn’t going to make some boring push-pram, requiring a mom or nanny to use their own muscles to cause the thing to move, like some sort of filthy ox, but instead would build motorized prams.

They started fairly conservatively, with a pram that had a one-horsepower two-stroke motor, mounted horizontally under the pusher/driver’s feet, and driving its own fifth wheel:

One horsepower you would think would be plenty for a perambulator. Hell, my Changli only has 0.1 hp more, and it’s pretty close to a whole car! Sorta!

But remember, this Dunkley we’re talking about here. They DGAF. They knew that they could make a faster pram, and that was good enough for them, consequences be damned. In 1922, they showed their next motorized pram, powered by a 750cc single-cylinder engine making 21 hp!

Holy crap, right? That’s a bigger engine than the 603cc flat-twin in my Citroën 2CV and makes almost as much power! As you can see from that picture of the Duke of York inspecting the thing, they also gave it what is essentially a full metal automobile body. This is a car. It’s a car, for a baby, with an adult hanging on to those handlebars behind it. I can’t find any references stating how fast this thing could go, but I’m pretty sure whatever it was it was way, way too fast.

Oh, Dunkley! You were too crazy to live, if we’re honest. Gas-thieving cars and superfast motor-prams are just more than this staid world was able to bear, I’m afraid, but I happily salute you.

The post There Was Once A Carmaker That Designed A Car To Steal Gas (They Also Made A 21 HP Baby Stroller) appeared first on The Autopian.

The “hockey stick graph” is a graph which shows slow initial growth, followed by a rapid linear increase. The resulting shape resembles a hockey stick, with the blade of the stick represented by the nearly-flat initial section, and the handle of the stick represented by the rapid linear increase once sales take off.

All sales forecasts have this hockey stick shape because the people who do sales forecasts are all optimists.

The Microsoft finance division has their own variation on the hockey stick: The hockey stick on wheels.

Consider a team which presents their forecasts in the form of a hockey stick graph. They come back the next year with their revised forecasts, and they are the same as last year’s forecast, just delayed one year. If you overlay this revised hockey stick forecast on top of the previous year’s forecast, it looks like what happened is that the hockey stick slid forward one year. When this happens, the finance people jokingly call it a “hockey stick on wheels” because it looks like somebody bolted wheels onto the bottom of the hockey stick graph and is just rolling it forward by one year each year.

Net profit, net profit.
I love ya, net profit.
You’re always a year away.

An example of a hockey stick on wheels is the first few years of the infamous Itanium sales forecast chart. Notice that the first four lines are basically the same, just shifted forward by one year. It is only at the fifth year that the shape of the line changes.

The post Microspeak: The hockey stick on wheels appeared first on The Old New Thing.