I don't know, and I think there is no easy answer. The point is: the investors don't know how to measure traction either, so they just measure GitHub activity instead, even at the very moment in which it becomes obvious that it does not capture actual traction.
The absurdity lies in the statement that the developers still need to gain actual traction while putting additional effort into gaming that metric to satisfy their investors.
> It's especially sensitive for a VC-backed startup that is measured thoroughly by GitHub activity, but we have to pull the trigger:
This sentence also illustrates the absurdity of this investment model. It imposes a trade-off between building good software, and complying with the investor's metrics. They probably call such metrics evidence-based, but this example shows that they arbitrarily capture some numbers to obscure the lack of meaningful measurements.
In literally must have missed that. When did Microsoft ever encourage energy saving?
Is this related to power saving for extending laptop battery runtime? But then I don't get the link to renewable energy.
Anyway, I agree with the notion of the extreme energy-inefficiency of LLMs. The scale of it makes it hard to imagine any less efficient product will ever be invented.
You could abstract speech or other audio as a series of sounds, where time is indeed a factor. Speech, however, has patterns that are more similar to written language than to seasonal patterns that are typically assumed in time series.
While trained on different data, the architecture of TimesFM is actually similar to LLMs. But not identical, as pointed out at https://research.google/blog/a-decoder-only-foundation-model...:
> Firstly, we need a multilayer perceptron block with residual connections to convert a patch of time-series into a token that can be input to the transformer layers along with positional encodings (PE).
> [...]
> Secondly, at the other end, an output token from the stacked transformer can be used to predict a longer length of subsequent time-points than the input patch length, i.e., the output patch length can be larger than the input patch length.
> On top of that, I exported my location timeline from Google Maps, my Uber trips, my bank transactions, and Shazam history. I would ask Claude Code to start with the photos and then gradually give it access to the different data exports.
Is anyone else feeling uncomfortable with that? It is a great project and I don't want to bash it with general concerns, but sharing all my financial and location details with any service seems like opening the floodgates to my house.
My concern is not even strictly related to AI, but about sharing all my most private data with any service. There is always a significant chance all of it is leaked sooner or later.
On the one hand, this exercise probably reflects a realistic task. Daily engineering work comprises a lot of reverse engineering and debugging of messy code.
On the other hand, this does not seem very suitable as an isolated assignment. The lack of code base-specific context has a lot of potential for frustration. I wonder what they really tested on the candidates, and whether this was what they wanted to filter for.
Essentially, the LNM turns out to be unsuitable for various diagnostics it has been applied for:
> Our findings reveal a foundational limitation: at its core, LNM involves a repetitive sampling of one and the same FC matrix. As a result, it systematically maps sets of local brain changes—whether they are patient lesions, magnetic resonance imaging-derived alterations, synthetic or random—onto the same nonspecific properties of the used FC data, producing highly similar networks across conditions.
I suppose you are right about the history of firearms. However, the novel was written in 1844, more than 200 years after the time in which it is set.
Which makes me wonder if the author (Alexandre Dumas) knew and cared about the historic facts.
It is concerning that GitHub hosts the majority of open-source software, while actively locking its users into a platform that is based on closed source for eerything except Git itself.
This issue with Actions shows how maintaining proprietary software inevitably ends up rather low on the priority list. Adding new features is much more marketable, just like for any other software product. Enshittification ensues.
Technically not OCR, but HTR (hand-written text/transcript recognition) is still difficult.
LLMs have increased accuracy, but their mistakes are very hard to identify because they just 'hallucinate' text they cannot digitize.
Sure, euphemisms are an essential component of any kind of propaganda. To some degree, this includes legitimate campaigning in democracies.
What the Nazis implemented to a unique extent was the process of (re-)defining words in a specific meaning. This happened top-down and censorship not only acted passively (disallow certain publications), but the propaganda actively pushed the terminology and narratives to use by the media.
The recurrent use of seemingly harmless words in specific ways impacted the society much longer than the Nazis were in charge. Some seemingly unpolitical narratives have lived on until today. Fortunately, the (Western) allies realized that they needed to counter this after the war, but of course the denazification has not been able to undo all the damage.
> LTI demonstrates changes in the German language in most of the population. In contrast, the text also emphasizes the idea that resistance to oppression begins by questioning the constant use of buzzwords.
While supporting all the individual tips here as well, trying to understand the underlying societal dynamics of mental issues can also create new perspectives, new motivation, and relieve.