exabyte scale data strategy by day job. safety critical aviation/marine systems by passion. Not involved in crypto; short for Corvus cryptoleucus :D

Stockholm, Sweden
Joined September 2013
I'm lost, do people think tracking systems today convert a tracked object only into coordinates sent to the cloud for GPS tracking? This was the standard long before GPS. Seek, acquire target then track. Payload detonation optional. The actually cool thing here is the obstacle avoidance, which had nothing to do with GPS even before this, but is a cool thing to solve
Purdue University's IDEAS Lab built an autonomous drone that tracks runners through dense forests completely without GPS or cloud networks. By running 3D LiDAR mapping and obstacle-avoidance AI entirely on-chip, it can weave under tree canopies and predict target movement through blind spots in real time. Massive implications for next-gen search & rescue and edge AI.
39
Incredible momentum in betterment we are seeing. github.com/alrojo/fp4Attn
🥐 Hi everyone! FlashAttention's backward pass has the same zero-sum bug @jon_barron just fixed in softmax. With FP4 stochastic rounding it's wrong on every single draw. We fixed it in our ICLR'27 submission, CroissAttn and unlock e2e FP4 training.
2
108
Building a Diffusion Transformer model has proven to be more about scheduling what conditionals you introduce to steer more than just learning structure from the entire set. This has become annoying, but client pays so work continues.
22
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
34
This is a great shot across the bow for all others to do at least 3 things: 1) evaluate your existing technical risks and have a continuity plan for even "unprecedented" outages. You don't need to build for every miniscule probability of failure but at least think about the scenarios. 2) review your comms procedures for incidents and prepare your teams for writing proper post-mortems. Have a standard way to communicate what is being done and ensure you are prepared to do so at regular intervals. Include support for workarounds that help during mitigation. 3) remind yourself that your customers care about YOU, not your providers. Everything that leads to customer disruptions is your responsibility to address and no one accepts finger pointing no matter how much weight the other party adds to an incident. These are all things Revolut failed to do here that will absolutely cause them to lose customers. Don't be this unprepared.
Due to a recent Google Cloud outage, customers may have experienced temporary service disruption on the Revolut app today. The issue has now been fixed, and services are now restored. All our customers’ money and data is safe, and has remained safe throughout this incident. We apologise for any inconvenience caused.
77
least broken html renderer on Arxiv btw
1
3
204
I can't stress how important it is to speak the local language when you are in transport. During a switch outage for trams the driver for a replacement route couldn't understand the questions from older folks that can't speak english. I helped as I could but that is shameful.
1
40
And yes the provider has the feedback in their official channels but this is so bad to see.
12
Traveled to Barca for some unrelated business but this came up while talking with a larger Fintech firm. The stuff people are doing with agents is not the rosy picture most on X are painting. It's rough and many more companies will start doing this. Personally i am expecting the larger schemes to start instituting policies as well for blocking authorization of any Agentic transaction due to the high abuse and risk to consumers
If you didn’t see this coming , you must be incredibly naive.
1
84
Cool, but this is not the hard part about flight controllers. Not even close. It's the actual hardware testing and tuning for that which is hard. We already know what the boards do and how to test them easily for simple control loops. And don't get me started on how many of these things assume immediate power change on motors and ignore rotor dynamics completely
1
81
A lot of ML assumptions you make when building a model are based on what you believe are the right factors that condition the probability space you want to work in. And I will be honest, we get it wrong a lot. Like a lot a lot. This is why things like RLHF layers sprinkled on top plus gated MoEs are so heavily used. Much easier than having to go back and rearchitect an entire foundational model architecture based on new ways to condition the core blocks within the model (and cheaper ofc)
2
1
90
This is also obvious to many that work with other statistics models to be fair, but I fear many do not actually give a shit about how statistics and probability works and how it fits into the conversation of a machine that minimizes deviation from a distribution of data
20
Technically generality is where models push things anyway given ML is often about iterative minimization of error w.r.t. a reference distribution of data. 🤓 regression to the mean is something we've been seeing for a while now in most applications of many different classes of general models. I think it's natural to call things generic when certain styles or things feel ubiquitous and it will probably mean more effort long term with AI tooling to differentiate a good product from the res.t don't know really though because on the other hand people are sometimes okay with generic stuff. I'm just here to watch and enjoy the ride
I wish experienced developers would stop shitting on people who build things with AI. “AI slop.” “Generic.” “Broken on mobile.” A person that doesn't know the first thing about programming, but had an idea, built an actual product, and had the courage to share it with the world, gets bullied basically. How about a little encouragement instead? Everyone starts somewhere!
48
På tal om hur EU trampar på oss som är en del av skattebetalare i Sverige och är tvungna att leva mellan två världar... Jag skulle vara mycket tacksam om du kunde hitta någon att kolla (på riktigt) det här problemet. xitter.cf/corvuscrypto/status/21…
Svenska skattebetalare är ingen bankomat åt EU. Det är mitt budskap här i Bryssel idag.
59
Replying to @vonderleyen
@vonderleyen new EU local-branch banking rule further pushes toward a dual class system where those that have accounts outside of Europe are forced to sacrifice wealth held in other accounts due to the poor implementation of the new banking rules. Most do not adhere to the grandfather rule for fears of penalty. Any way to get some eyes on this to ease the pain? This is a bad implementation that disproportionately affects low-middle income individuals that must maintain accounts abroad for practical reasons. For those of us tied to the US this is like being punished twice, first with FATCA pushing European banks to limit access to services. And now this. If the goal is to push back on the US maybe don't punish those already caught between the fight unwillingly.
2
1
1
854
This is an example email and I have confirmed that it applies in Chase at least to all balance-holding accounts despite when they were created or how much is there. I would say it could be useful even if there is just a wealth cap. That lessens the blow significantly. I have redacted the listed accounts for privacy.
1
79
We were prosed reasonable exceptions and they are not being applied in a way that banks feel comfortable enough to allow existence of accounts that would otherwise be safe to maintain. This is the core issue. The amount of money up for loss here is, to me, significant. Plus the lack of access to investment vehicles that every other resident has access to. Financial segregation in other words.
46
Cc @jonasfernandez while this bill may have good intentions it misses that I already am limited due to FATCA compliance causing me to for example be unable to invest in assets in European banks that decide to slip that responsibility. Happy to discuss the practical realities here.
1
153