eamsen

Anecdote: Gemini 3.5 casually added a DROP TABLE for an actual production table in a system test.

It had previously attempted to create that table as part of the test setup, so it apparently concluded that it was a test table.

During human review, it explained that it had simply chosen a table name inspired by the codebase.

This is obviously not an NVIDIA site, I don't know the legal repercussions for such fake sites, but apparently it's worth the risk?

Our family derives a lot of joy from stylized versions of our photos. For us, it is not a waste. If you do not derive anything positive from it, you will likely not use it, hence no energy wasted either. Your argument is objectively wrong.

Personal anecdote: I've asked Gemini 3 Pro to write a test for a function that depends on external DB data. It wrote a test that creates and deletes a table, it conveniently picked the exact production table name, didn't mock the DB interactions. Attempted to run the test immediately.

We've entered the sand worshiping era. Except this time, we're the gods.

You are being printed, every day, cell by cell.

One distant day they’ll wonder who began them. Do you sign the work?

Can you please provide examples for the raw data? As a user, I would like to know what I'm buying before paying.

Completely agree with this statement.

I would go further, and say we don't understand how next-token predictors work either. We understand the model structure, just as we do with the brain, but we don't have a complete map of the execution patterns, just as we do not with the brain.

Predicting the next token can be as trivial as a statistical lookup or as complex as executing a learned reasoning function.

My intuition suggests that my internal reasoning is not based on token sequences, but it would be impossible to convey the results of my reasoning without constructing a sequence of tokens for communication.

A healthy society produces surplus to provide for those who depend on others. To ensure enough surplus, everyone who is able, should add value.

Value production is needed, because the value we produce is fleeting and healthy societies grow.

How much value is needed is determined by the society through a free market.

To add value to society?

Have there been catastrophic failures of fission reactors in Germany built by Siemens?

Are you familiar with functional safety classification and how this affects the development of systems? Would you draw conclusions from the failure rate of Siemens vacuum cleaners to the company's capability of building fission reactors?

Build: Siemens, AEG.

Operate: PreussenElektra, EnBW, Vattenfall, RWE, EWN.

Or was that a rhetorical question for a solution with a zero catastrophic failure rate (in Germany)?

Yet, here we are, generating text. Some of which has value.

It would depend on whether she got the job for having a good voice or for having a voice that is associated with a famous person. Would her voice have the same value if it was not sounding like another famous person's voice? That's up to the courts to decide.

At the risk of sounding snarky, https://chat.openai.com would be a good introduction, followed by books, which GPT could recommend.

FWIW, the loop doesn't seem to be continuous, eventually the page loads. Not sure if relevant, but I switched focus away from the tab during loading.

How much are you willing to pay for a browser?

No, there is no support for third-party shared libraries on Android yet.

> There are ZERO settings exposed to the end user.

This thread is discussing GeckoView, which exposes extensive privacy settings.

There is a (hopefully growing) amount of apps based on GeckoView, which will make use and/or expose different variations of GeckoView settings to the user, depending on the purpose and target audience of the app.

You might disagree, but, for example, I consider the level of detail in privacy settings provided by Fenix (Nightly) to be a good balance between user control and comprehensibility.

> > GeckoView also exposes a WebExtension API > again to the app.

Again, taking Fenix (Nightly) as an example, the GeckoView WebExtension API is used to provide a growing selection of add-ons, many of which are aimed at the privacy-conscious user, including uBlock Origin, HTTPS Everywhere, NoScript and Privacy Badger.

> For example, any new tech introduced that can be used for browser fingerprinting and user tracking (e.g. canvas, webgl, webrtc, etc) is enabled to every single site by default, just like chrome.

GeckoView exposes a wide range of privacy (including anti-tracking and anti-fingerprinting) features in a (hopefully) comprehensive API, see https://mozilla.github.io/geckoview/javadoc/mozilla-central/..., and continues expanding its privacy settings.

> Some you can install extensions (you can't on this and on mozilla focus)...

GeckoView also exposes a WebExtension API, see https://mozilla.github.io/geckoview/javadoc/mozilla-central/..., and continues to expand extension support.

> ...some you can fiddle with user-UNfriendly settings in about:config...

You seem to have answered one of your own questions here.

I also have the impression that you might have not commented on the GeckoView library, but on some range of Mozilla products (you named Focus).

Relevant question with answers: https://stackoverflow.com/questions/61375629/allow-cross-ori...

You're correct about that there are different degrees of seeking support.

Seeking VBR MP3 with perfect accuracy is trivial by forward-reading. However, this is obviously highly inefficient on long duration seeks.

For instant seeking support, you need to (partly) depend on the optional VBR headers. This comes with its own set of issues, e.g., the most commonly used Xing header contains only 100 seek table entries, which may not provide enough resolution for large files.

I'm still surprised about the complete lack of support for those headers in AVFoundation, since I would consider it a low-hanging fruit in terms of improving usability for the majority of use cases (excluding pod casts).

Disclaimer: I've worked on MP3TrackDemuxer for Gecko/Firefox.

You'll find the e10s roadmap here: https://wiki.mozilla.org/Electrolysis#Schedule_and_Status

The plan is to make it the default in Firefox 51 on desktop, assuming the staged roll out is successful.

Also, the current approach will not considerably increase resource usage for heavy tab users as FF will only run one content process shared across all tabs.

Interesting, I've never had to sign an NDA before my interviews (at top tech companies) including the on-site/offer stages.

I don't share questions as a matter of principle, fwiw.

Firefox still holds over 30% (#1) market share in some select markets, e.g., in Germany: http://gs.statcounter.com/#browser-DE-monthly-201504-201604

Xing is a popular alternative for the German-speaking markets.

The robustness experiments are highly parallelized and maintain a copy of the (modified) network per thread. I vaguely remember the NYC test would take around a day on 16? cores and use up to 64GB of RAM.

You can find query time evaluation in the performance recap on the results page I linked. For NYC it's around 2.2s for the Dijkstra (baseline) and 27ms for the TP-based search.

For single-threaded pre-computation and shortest-path queries, I would expect you to need around 8GB for NYC, less for Toronto and you can get the Honolulu feeds to run on <2GB (which was my local test set).

Sorry for not being more specific or inaccurate.

You can find some GTFS feeds here: http://www.gtfs-data-exchange.com

For anyone who's curious about an example implementation of transfer patterns: https://github.com/eamsen/transit-planner

Transfer patterns also behave robustly given real-time updates on the network, which we have shown here: http://stromboli.informatik.uni-freiburg.de/student-projects...

On top of being an excellent researcher, Hannah Bast is one of the best professors I've had pleasure of studying under and working with, so I'm happy to see her getting credit for this.