Fascinating research about side-channel attacks. Learned a lot about them and website fingerprinting here. Also interesting the explanations of how the use of machine learning models can actually get in the way of proper understanding of the side-channel really used by an attack which can prevent developing actually useful counter-measures.

Very interesting research. Looks like we're slowly moving away from the "language and thinking are intertwined" hypothesis. This is probably the last straw for Chomsky's theory of language. It served us well but neuroscience points that it's time to leave it behind.

Now this is an interesting paper. Neurosymbolic approaches are starting to go somewhere now. This is definitely helped by the NLP abilities of LLMs (which should be used only for that). The natural language to Prolog idea makes sense, now it needs to be more reliable. I'd be curious to know how many times the multiple-try path is exercised (the paper doesn't quite focus on that). More research is required obviously.

Now the impact seems clear and this is mostly bad news. This reduces the production of public knowledge so everyone looses. Ironically it also means less public knowledge available to train new models. At some point their only venue to fine tune their models will be user profiling which will be private... I've a hard time seeing how we won't end up stuck with another surveillance apparatus providing access to models running on outdated knowledge. This will lock so many behaviors and decisions in place.

Of course I recommend reading the actual research paper. This article is a good summary of the consequences though. LLMs definitely can't be trusted with formal reasoning including basic maths. This is a flaw in the way they are built, the bath forward is likely merging symbolic and sub-symbolic approaches.

Still very early days on this topic, clearly more studies are required. Still this one is interesting and indicates are clear link between code review anxiety and code review avoidance. If you're often procrastinating or rubber stamping code reviews, a workshop to reduce biases and showing you can manage your anxiety could improve things greatly.

OK, this paper picked my curiosity. The limitations of the experiments makes me wonder if some threshold effects aren't ignored. Still this is a good indication that the question is worth pursuing further.

This is a short article summarizing a research paper at the surface level. It is clearly the last nail in the coffin for the generative AI grand marketing claims. Of course, I recommend reading the actual research paper (link at the end) but if you prefer this very short form, here it is. It's clearly time to go back to the initial goals of the AI field: understanding cognition. The latest industrial trends tend to confuse too much the map with the territory.

I'm obviously not in love with the complexity this type of architecture brings. That being said, this thesis brings an interesting approach to better detect failure scenarios in such systems.

It's good to see major institutions like this get out of contracts with scientific publishing companies. Those unfortunately became mostly parasitic. Open access should be the norm for research.

More discussion about models collapse. The provenance of data will become a crucial factor to our ability to train further models.

Further clues that transformer models can't learn logic from data.

Interesting paper showing a promising path to reduce the memory and workload of transformer models. This is much more interesting than the race to the gigantic size.

Another cruel reminder that basic reasoning is not to be expected from LLMs. Here is a quote from the conclusion of the paper which makes it clear:
"We think that observations made in our study should serve as strong reminder that current SOTA
LLMs are not capable of sound, consistent reasoning, as shown here by their breakdown on even such a simple task as the presented AIW problem, and enabling such reasoning is still subject of basic research. This should be also a strong warning against overblown claims for such models beyond being basic research artifacts to serve as problem solvers in various real world settings, which are often made by different commercial entities in attempt to position their models as a strong mature product for end-users. [...] Observed breakdown of basic reasoning capabilities, coupled with such public claims (which are also based on standardized benchmarks), present an inherent safety problem. Models with insufficient basic reasoning are inherently unsafe, as they will produce wrong decisions in various important scenarios that do require intact reasoning."

Nice article. It's a good reminder that the benchmarks used to evaluate generative AI systems have many caveats.

This is how it should be done. This one comes with everything needed to reproduce the results. This is necessary to gain insights into how such models work internally.

More work about eco-design of software. This is definitely welcome. I found this work a bit weak on the state of the art and the interview parts (10 people in the same company). But the field is so nascent that it's to be expected I guess, PhD students have to do with what they have access to. Unsurprisingly this shows a great lack of proper tools to tackle the measurement problem. This thesis shows interesting prospects to reduce variations in measurements though, some of the proposed guidelines might help but cannot offset the hardware heterogeneity completely... The parts focusing on practical advices around Java use and deployment are interestingly easy to apply though. You need to take into account the context of your application to make the right choices of course.

This is great news, more scientific papers from the past decades will be accessible to everyone.

An important question for proper statistics about the content itself. Surprisingly harder to get an answer to it than one would think.

Interesting research, this shows opportunities to push CRDTs to the next level.