alex_hirner

A combined compression and indexing scheme that actually has code is Succinct by Berkeley [0]. I remember being impressed by the tradeoffs you can dial in and its performance in a 2015 presentation [1].

[0] http://succinct.cs.berkeley.edu/wp/wordpress/ [1] https://www.youtube.com/watch?v=R_SHYey7eas (Video)

Rest assured they work with a company that is really good at packaging [1]. However true, chemistry innovations will need to come from the lower end of the R&D pipeline.

[1] wrote my thesis with them, http://openbatt.org

That's right and why I used it (v2) extensively.

I think the implications go even beyond privacy and efficiency. One could estimate each user's contribution to fidelity gains of the model. At least as an average within a batch. I imagine such an attribution to rewarded in money or credibility in the future.

I think the trick would be to suggest suitable pictures per paragraph based on NLP. If to take the highest rated content, there would already be a (biased) learning set available.

I would pay a few bucks to have my articles proof read and for stylistic enhancements. How much this producer centric model would yield is another question...

The model clearly overfits if to compare the fidelity of pictures from the training set vs the out of sample ones (sent in by others).

I haven't but the ressources on jetsonhacks.com are amazing to see robot vision on this board in action.

Inference times for inception-v3 on a Raspberry Pi have been benchmarked recently [1]. They weigh in at around 2s. If you remove the python layer it gets down to 500ms. However, it's not yet clear if the python overhead is the only reason for that.

Nvidia's Tegra X1 should supposedly be capable of <10ms for imagenet grade models [2]. It's fair to assume though, that this must be for trimmed down and/or 16bit models as compared to full inception models.

And finally, Sam who also facilitated building TF on the Pi is about to host a 6 weeks half theory, half practice course on TF and deep learning [3] (me thinks he deserves this plug).

[1] https://github.com/samjabrahams/tensorflow-on-raspberry-pi/t...

[2] https://youtu.be/_4tzlXPQWb8?t=53m35s

[3] https://www.thisismetis.com/deep-learning-with-tensorflow

I wonder which "SLAM for the masses" technology will become dominant. Do you have any thoughts what's missing in google's project Tango?

> Loosely speaking it's like how the math that physics uses doesn't have mutable variables

I'm curious, can you elaborate on this? If you view time as discrete dimension, mutable state is clearly not necessary. But since we cannot travel through time (yet), nature seems quite mutable in practice. For example, when microwaving food, it is extremely hard to put it back to the same temperature gradients.

Similarly with functional programming: if you adhere to it purely, provenance of every state can be traced back. If not, state changes are non-trivial to reverse. So it seems like modelling physical phenomena with immutable states is a choice of perspective rather than a law of physics that can be empirically tested.

No. The trust you need for a central curator is trivial compared to what's needed toward fundees. Anonymity aggravates problems with accountability. Thus, I don't find crowdsourcing to be a particular good example where blockchain-ideas help to create and coordinate collective action. Yet, I think there are two areas where value can be added to kickstarter: 1) scheduled payouts where funders vote to cut project funds in absence of progress and 2) make the curation process more transparent.

However, I figured that what's holding all the blockchain collaboration platforms back are not only scalability and cost issues. More importantly, people don't want to broadcast their data for very uncertain returns. You can see that in haggling parties for when and who gets access to aggregate data in the connected car, health care and scientific research.

Therefore, I focus to find meaningful ways around data-sharing tensions, specifically in battery R&D [1]. One way is to align academic and commercial interests better. The other way is to objectively measure what a data donor's synergistic effect was on the platform as a whole. This kind of computable attribution is hard, but I think doable at some point and in some narrow domains earlier.

Some infrastructure that allows to selectively reveal and collaborate on data is put forward by guys at MIT [2].

[1] http://openbatt.org [2] http://www.enigma.co

Talking about using "the blockchain" doesn't qualify. Rather the opposite. The real value will be on the fringes of that ecosystem and has little or nothing to do with chaining blocks: secure multiparty computation, testing the potential of synergies in data with bloom filters and differential privacy.

Source: running an ethereum meetup since 2014 and (very likely) writing the world's first crowdfunding contract in LLL.

The decision makers in big auto are around 10y from their retirement. Reluctance makes sense for them. It's similar to the consulting project I had at uni with an O&G company: everyone knows they should pivot for the sake of the next generation. On the other hand, upper echelons just want to simplify their life.

Now it's the labor unions that are pushing german auto makers over the EV cliff. Ironically, they have the greatest incentives to innovate/follow.

It's bothering though, that redundancies will be massive no matter what exactly they do about EVs. Can anyone offer a growth perspective for that sector (=not some kind of UBI vision)? My initial thoughts were space. But this is not really were to retool design prowess and factories either.

Funding bodies emphasize data sharing more and more. However, that is along or after a publication.

No one has figured out academic attribution of data-only in a satisfactory way yet, but some noteworthy attempts exist. E.g. https://www.elsevier.com/connect/new-data-journal-lets-resea...

Not to forget chronix.io, which was shown to have favorable performance for more complex queries and space efficiency at last week's http://munich-datageeks.de/dadada/ (slides yet to come).

In order to transfer learnings from virtual worlds to the real world more efficiently, I imagine a module that translates real world footed to an abstract representation that is the same as for rendered footage.

In other words, you would train a NN on input from a footage of a walkthrough of a real building to output rendered footage from the same path (3D data from architects is already there). The middle layer of this quasi autoencoder then is the basis to train fully simulated tasks, e.g. autonomous vehicles. In a way, it would be similar to colorization of b/w footage. Would that scale training data?

An E-2 investor visa could also do the trick in case of affluent resources. I wouldn't know about limits on the controlling stake though.

> Did it mean that 50% of YC batch applicants for EB-1 got it?

That was unclear of me. Actually, the ratio of granted vs. rejected applications dropped by around 50% (from a discussion in Feb this year). Thus, I don't know how many applied for in total in a typical cohort.

> I am trying for EB-1 but even with my background it will be difficult.

If you have probed this possibility already, what is the evidence that speaks against you reaching EB-1A status? I assume making three out of the ten categories won't be problematic, but I myself am unclear about qualitative judgments in totality.

What I do know is, that the many EB-1 grantees from any YC batch were decimated to 50% sometime around 2014.

I have not looked at the filings, but R&D expenses are usually amortized much like factory equipment. At least, activating R&D as an asset is often exercised in the IFRS world.

Nevertheless, ramping up R&D disproportionally to current sales like Tesla does also incurs disproportionate costs compared to rather static incumbents.

PS: The high SG&A figures are still confounding me. In other sources, Tesla's marketing budget was said to be negligible for the premium segment.

What a beautiful project. I started to hack around for a guitar amp emulator on the bb in 2014, but it quickly became too involved. Thanks for sharing.

If there is one thing you can change in the process of battery R&D, what would it be?

And likely a related question: Which scientific data is not readily available but would help battery R&D?

Where can the reviews be read and do they involve other vendors too?

All sorts of problems with lithium-metal as we know it. A more benign version pitched by ZAF Energy is zinc-air. They inherit a few problems from pure metal electrodes (like dendrite growth) but not its thermal volatility. On the flipside, the performance increase will practically be not more than 2x that of lithium-ion.

ZAF had a very in-depth visit at Microsoft Research in June 2015. It is interesting to see the tribulations of a startup in commercializing technology that might become feasible but isn't quite yet and maybe never will. The 1h+ pre-sale meeting was taped [0] and I have an article in the writing about it [1].

[0] https://www.youtube.com/watch?v=xWwvHPaimlU

[1] https://docs.google.com/document/d/1R5gkh_A2vb6mDI5-e-EBlvP8...

You must be talking about the flow version of lithium-sulfur. The ones with (presumably) graphite and cobalt oxide electrodes are going into e-scooters in China now [0]. However, they are yet far from the theoretical 2x in volumetric density of lithium-ion.

For a very good survey on the state of the art in lithium-sulfur look up this recent study [1].

[0] https://twitter.com/openbatt/status/766641735229464576

[1] http://doi.org/10.1039/C5CS00410A

Differential Privacy has its limits for small data sets. Luckily, more research is on the way to run arbitrary calculations on multiple datasets that remain private in its entirety (basically homomorphic encryption, but efficiently) [1]. Algorithms that work reasonably well are simple statistical tests and ridge regression [2].

I think these innovations are rare instances where technology can really solves a social problem.

[1] https://www.microsoft.com/en-us/research/microsoft-researche... [2] https://www.microsoft.com/en-us/research/wp-content/uploads/...

The argument is reversed entirely if you take human desires as given, i.e. if you acknowledge that people do text on the road.

The metric to compare both alternatives then becomes, how many miles of texting while driving vs. being driven by an autopilot while texting does it take to kill one person. I don't have the numbers, but I'm pretty sure the latter is lower. Maybe autopilots encourage "bad" behavior, but convenience also spurred construction of coal power plants at the cost of shortened lives. Much like fines for neglected pollution abatement, I would favor mandatory life insurance for each mile on autopilot to internalize risks.

In contrast to popular fiction, my concerns don't lie in a conspiring super AI. Rather, we should take countermeasures against suffering from a thousand cuts. Flash crashes in financial markets are a precursor to other sytemic risks ahead. Maybe this is a distinction that helps to appreciate safety for AI more.

Bigchaindb is tackling generic open data repositories from a technical pov. As has been said below, the larger part of the sustenance prob is economical of course.

There is significant effort (and progress) in standardizing and contrasting test methods [1]. This usually constitues a good first step in a market of lemons. [1] http://energy.gov/sites/prod/files/2014/09/f18/es_cunningham...