In a blog post published Thursday, the company cited its own internal data as evidence that modern AI systems are nearing the point of “recursive self-improvement”—i.e., being able to refine their capabilities without a human in the loop. “AI that can build itself would be a major development in the history of technology—one that could bring enormous good for the world in science, healthcare, and beyond,” the post, which was written by company cofounder Jack Clark and Anthropic Institute lead Marina Favaro, reads. “But full recursive self-improvement also might increase the risks of humans losing control over AI systems.”

Are they? When our own researchers barely have a grasp on how LLMs work, can’t really pinpoint why LLMs make certain choices, can you really expect it to improve itself when it has no sense of understanding? Or will it just make it self even more convoluted and nonsensical until it all falls apart like a house of cards?

I don’t know shit about LLMs, so maybe I’m lacking understanding here.

Another peculiarity discovered is that the operator hardcoded the recovery phrase of their own cryptocurrency wallet. The researchers say that the only reason for this is that the threat actor did not want the malware to steal it during the test stage.

Wait, what?

No

Maybe AI can play my games for me now, too?

Some people interpret the name as “no stupid questions allowed” for some reason.

We’re posting tabloids now?

The Lemon about page is rad as hell.

This is accomplished through what is known as WiFi sensing, or the use of WiFi signals to infer information about a physical environment. When radio signals like WiFi travel through a space, they interact with the objects and people around them. Those signals can be reflected, scattered, or absorbed. By analyzing how the signal is expected to behave compared with how it is actually received, researchers can infer details about the surrounding environment.

I kind of get this idea. It’s just sensing the strength of radio waves with some kind of antenna array. Like pixels on an optical sensor. It’s not really a new idea, I sort of recall seeing video of this kind of thing from like 10 years ago.

But the paper suggests this relies on higher level protocol information.

Beamforming, as introduced in WiFi 5, requires clients to broadcast observations of their channel characteristics. This introduces a new information source for WiFi sensing with privacy threats that have not been explored, so far. With WiFi networks being ubiquitous in our everyday lives, the impact of unknown privacy threats is likely severe. To investigate this concern, we introduce BFId, the first identity inference attack using BFI-based sensing and evaluate its efficacy on a novel dataset containing WiFi recordings of 197 individuals. We show that we can infer the identity of individuals with very high accuracy, across different walking styles and perspectives, even with large sample sizes.

[…]

Identity inference based on WiFi can be done by analyzing different sources, but most prominent in recent years has been the analysis of Channel State Information (CSI), a built-in part of the physical layer of WiFi. […] To enable higher bandwidths, WiFi 5 (802.11ac) introduced beamforming. Beamforming utilizes similar information on the physical environment as CSI, but on the sender instead of the receiver side.In a typical WiFi scenario, clients send Beamforming Feedback Information (BFI) back to the access point, a compressed representation of the current signal characteristics

This makes it sound kind of like a way to fingerprint signals, but they also mention:

We show that individuals can be recognized with very high accuracy (99.5% ± 0.38) with our BFI-based attack. Furthermore, BFId is not only able to infer the identity of individuals, our experiments also demonstrate that in a direct comparison it is able to do so better than CSI-based attacks for large populations. This also holds for identifying individuals across walking styles, from multiple different perspectives, and at reduced sample rates.

But if we’re talking about walking styles, it sounds more like “regular” wifi sensing again.

For this, the standard defines a channel sounding procedure (shown in Figure 1) which is initiated by the access point (beamformer) regularly through a null data packet (NDP) announcement frame. Beamformees will reply to this announcement. The actual NDP is then sent by the access point which contains one VHT-LTF (very high throughput long training field) per spatial stream used in the transmission. Beamformees will then use the CSI of these VHT-LTFs to calculate so-called feedback matrices for each subcarrier. The feedback matrix is compressed into beamforming angles which are sent back to the beamformer.The beamformer can then calculate a steering matrix which can be used to direct the transmission towards the beamformee [6, 16].

So I guess it’s kind of like sensing from multiple observers?

The vast majority of approaches use recordings of gait sequences for identification, but there are exceptions, e.g. using lip-motions [ 42], keystrokes [ 18 ] or no moving at all, but the individual just standing [53 ] or sitting [51 ]. When usinggait, most approaches record individuals while walking orthogonally to the line-of-sight (LOS) between sender and receiver. Twoearly approaches had participants walk parallel to the LOS [26, 71 ] and some approaches have opted to have participants walk freely,but only one approach considered multiple perspectives [76].

Yeah, I guess the novel thing in this paper is using information from this protocol to essentially get a bunch more sensors. Though I’m still shocked they can put together a coherent picture to get fucking lip movements with all this.

Also, WiFi has gotten so much more complicated since 802.11b. Feels like I forgot to pay attention.

undecipherable nonsense

That’s a big box.

EDIT: I like the high tech foil.

New consultant for the US.

Watching a beige box seems kind of dull.

Ever hear of trepanning? People do stupid shit for stupid reasons.

You have too much milk.

Manifold Garden was great.

I wanted to like that, but I remember it being janky as fuck.

I would also like to know what this means.

You can rub the kind ones, too.

I want to like smart glasses. And wearable tech in general. Too bad it’s all entirely centralized, monetized, and exploited for political gain.

I really want to gut my car’s infotainment system. I’m just worried about the resale value and whether or not part of this shit is somehow critical to the vehicle now.

Not like there’s any useful manuals on this shit out there. And the manufacturer is incentivized to actively thwart my efforts.