HNHacker News
TopNewBestAskShowJobs

bnjemian

404 karma · joined May 31, 2016

AI/quantum scientist, 0-1 software engineer.
submissionscomments
bnjemian··on The fall of Sam Bankman-Fried is crypto’s Enron moment
Can't say I care too much about FTX's woes. Nonetheless, this article has some janky moments. The author cites an unnamed friend "who is deeply involved in [the crypto] industry" when saying CZ played a master stroke against SBF. Get a real source or at least demonstrate that your "friend" has an opinion worth 2 cents.

Maybe CZ did outplay SBF, but aside from the optics it doesn't seem anywhere that salacious. As was summed up earlier today in an NPR segment on this story "airlines don't compete on safety." In other words, it is in the interest of every exchange that the exchange industry is on the whole viewed as being comprised of trustworthy actors. Crashing FTX doesn't seem an optimal move given the already precarious state of the crypto industry these days.

Speaking of optics, the author also made a nonsensical analogy "more opaque than a dog with glaucoma" when describing the inner workings of FTX. Quite certain they meant cataracts – the clouding of the lens of the eye – not glaucoma, which is damage to the optic nerve.

bnjemian··on The sinister attempts to ‘decolonise’ mathematics
That's a fair point – equity is a subtle one – but it can be both important in some cases and problematic in others.
bnjemian··on The sinister attempts to ‘decolonise’ mathematics
Not so long ago I was a PhD student in a STEM department. With the amount of DEI and antiracist discussions had during our department meetings, it often felt that I was attending an MBA program designed for HR professionals.

Do I have anything against diversity, equity, inclusion? No.

Did I often wish the department – which was already extremely diverse in age, gender, ethnicity, race, sexual orientation, and citizenship status – would focus more on the work of doing good science, facilitating rich academic discourse, fostering cross-institutional collaborations, and driving more translation and commercial ventures? Yes. 100%, yes.

In many ways, DEI initiatives, anti-racist policies, and decolonization aspire to worthy goals. Too often, though, they feel like the product of meme-driven corporate training and consulting firms who charge wholly unequipped professionals with disparate expertise with framing their work with a funhouse mirror of their on-the-ground reality.

There has to be a better way.

bnjemian··on Ask HN: Advice that changed your life?
I once took a golf lesson when I was a kid and, in learning about the divot removal tool, was told a similar one about the way to operate on a golf course:

"Leave it better than you found it."

I couldn't care less about golf, but this one stuck with me because it speaks to environmental stewardship. It's like a constructive riff on the mantra for when you're out in Nature: "Don't leave a trace."

bnjemian··on Lung cancer rates by state, 2017
Hadn't heard of this. Yikes – dirty sodas sound pretty vile.
bnjemian··on Lung cancer rates by state, 2017
Yup, and it's getting worse. This is largely driven by drought and the dessication of the Great Salt Lake. The dust pulled off the basin is also toxic due to heavy metals (most notably arsenic), industrial and agricultural runoff, and the remnants of algal blooms. Here's one recent article covering the topic: https://www.deseret.com/opinion/2022/6/29/23188005/opinion-t...
bnjemian··on Lung cancer rates by state, 2017
This is literally the answer. My brother's a data scientist at a large health insurance company and this is a subject we've discussed at length. In short, because of the large Mormon population in Utah, the state has shockingly good statistics on a very wide range of disease incidences, health-related quality of life outcome measures, etc. With respect to health and longevity, it pays not to drink, smoke, or do drugs (surprise, surprise).

Other interesting tidbit – his health insurance company (and I expect all others) took a substantial hit 2019-2020 from the initial phase of the pandemic. Come 2021-2022, large proportions of the unhealthiest segments of the population had passed away; covid + comorbidities puts you at a much greater risk of dying than covid alone. They integrated this fact into their modeling and, given that these people would no longer require care (e.g. dialysis, ER visits), predicted that utilization would go down substantially and, now providing for an overall healthier population, the company would be more profitable that year relative to pre-pandemic years. This turned out to be an accurate prediction.

bnjemian··on A Universal Cancer Treatment?
Not to be too pessimistic, but 600,000 people died of cancer in 2019, so 20-30 years is hardly "round the corner" when you extrapolate to the, let's say conservatively, 10 million families in the US losing loved ones in that time period.

The history of cancer (The Emperor of All Maladies is a good book covering the history) is full of promising adjuvants, drugs, protocols that fail to generalize well. There are a lot of reasons for this. With drugs, one is that Phase 3 clinical trials often have patients who are selected on the basis of them being likely to be among the best responders. But once the drug is approved and made available to all patients within a given indication – a fundamentally different population – an overwhelming positive response may be significantly more modest. In many cases, this has to do with a patient's tolerance of the side effects or the interaction of the drug with known or underlying comorbidities.

While I'm encouraged by the research in this article and could see the drug being part of a combination protocol, I'm hesitant that it will be "universal". And the mechanism, candidly, is downright scary – if I were running a Phase 3 for this (seems the trial cited was a Phase 2 demonstrating baseline safety in humans), I would want a very detailed articulation of how the technology ensures with a high margin of safety that the drug delivery is targeted to the tumor and no other tissues. The permeability of blood vessels in tumors would not be sufficient (and also does not seem "universal").

On a more personal note, I'm actually a computational biologist and have done quite a bit of work in cancer research (masters thesis, portion of my dissertation). My Mother was also diagnosed with a highly aggressive cancer of unknown primary (CUP) origin in her lung late last year (estimated stage IIIb for NSCLC, stage 4 for CUP/melanoma). Its location and heart involvement made it inoperable. Turned out is was a melanoma, which to be frank I quickly recognized down to the subtype upon reviewing the pathology reports. The lung oncologists were much more conservative; given the location, they weren't especially well-versed in melanoma, and presumed it was an adenocarcinoma with a rare presentation despite the staining for carcinomas being negative across the board. Luckily, we managed to convince them to split the difference on the standard of care – CarboTaxol (carboplatin + taxol) combination and immunotherapy (nivolumab and ipilimumab), all at once. The immunotherapy surely saved her life; unlike with carcinomas, chemo is roughly 5% effective (as in, any response whatsoever) for melanomas. About 70% of melanoma diagnoses respond to combination immunotherapy with about 15% going into full remission (memory is shaky on that last number, may be slightly higher). With >95 PDL-1 expression, she was one of the lucky ones – she had a full response on both imaging and pathological endpoints. That also made her tumor (or what was left of it) operable. She's two lobes and a chunk of heart lighter, but she's alive, healthy, and recovering well. And while I cringed when he did it, one of her oncologists dropped the "c word" after surgery in discussing her case. Knowing the foe, I'm much more cautious in contemplating whether any of this represents a cure. Psychologically, the uncertainty around the future maintains a heavy burden over her and our family. A 70% response rate sounds really good, but it's a very different calculus when you're living it.

But as you said, her path towards a cure wouldn't have been possible not too long ago – in her case, 10-15 years; ipilimumab was approved in 2011 and nivolumab in 2014. But 30% of people with melanoma are still non-responders to combination immunotherapy. And, while a handful of other (generally less effective) options exist, non-responses in melanoma are deadly, often within a year or two of diagnosis, and for most all cancers have an incredibly high opportunity cost.

bnjemian··on Our long, vulnerable childhoods may be the key to our success
This is hard to read. Seriously, both the author of the review and of the book (assuming the reviewer accurately portrays it) seem confused about evolution and have an incomplete or total lack of comprehension of ideas; like fitness landscapes, that evolution is not the result of a design process, and that traits can persist despite being neither adaptive nor maladaptive. Regarding the latter, there's a concept in cancer genomics – passenger mutations – which provides a DNA level example of this on a much faster timescale.
bnjemian··on Meat made from recycled carbon dioxide
It's not like it gets you nothing; photosynthesis would be much less efficient in terms of the land footprint required. The article discusses the footprint required for their facilities relative to normal agriculture; the difference is massive. Photosynthesis is effectively driven by a 2D flux and thus is constrained by surface area to a pretty large extent – unless you use an electric light source, which would require renewable energy to be sustainable and thus, again, be constrained by surface area (think wind farms and PVs). With their technology, they can use a bioreactor and aerate it with the CO2, etc. required to grow the microbes they're using. That means the energy is not constrained to a 2D flux and can be delivered across all three dimensions in a volume.
bnjemian··on Jailed baby gene scientist ‘should be made responsible’ for the children
The comment above by verisimi was factually incorrect. The Pfizer vaccine contains MRNA. MRNAs are short-lived nucleic acids (not DNA) that exert their function in the cytoplasm of the cell (not nucleus, where DNA lives and Cas9 exerts its' action) and are used to translate proteins. The proteins coded for by the MRNA sequences in vaccines (Pfizer, Moderna) code for the S1 spike protein, which occurs on the surface of the SARS-CoV-2. After their translation, these proteins are then recognized as foreign by the immune system of the host (i.e. person), leading to an immune response. There is no Cas9 enzyme in MRNA vaccines.
bnjemian··on How Quantum Computers Will Correct Their Errors
The title of this post is a bit misleading (the important 'How' is also missing from this posts title). Control of current (and any near term) QC is classical, including the control of any error correcting code, such as Shor's.
bnjemian··on IBM unveils 127-qubit quantum processor
There are a lot of languages. I use Python. Right now I'm working on a project using the Pennylane library for quantum machine learning. Other languages exist – dedicated ones like Q# or quipper (built on top of Haskell) or libraries in Julia, C++, etc.

One paradigm, variational quantum algorithms (part of the broader class of hybrid quantum-classical algorithms), has several similarities to classical deep learning. The approach can be summed up as:

1. Construct a parameterized quantum circuit – one where the quantum gates are controlled by real-valued parameters. Ideally this circuit is one you think could reasonably solve some problem based off of your knowledge of the problem, input data, and behavior of quantum information.

2. Define an objective function.

3. Iteratively adjust the circuit parameters. This is done by measuring the output of the circuit and using that information in conjunction with some optimization function (e.g. stochastic gradient descent).

In general terms, I'm working on QML algorithms that may some day be used for applications in biology and medicine.

For a serious discussion of the approach, see: https://www.nature.com/articles/s42254-021-00348-9

There's also several nice walkthroughs of the basics on Pennylane's website: https://pennylane.ai/qml/demos/tutorial_variational_classifi...

bnjemian··on IBM unveils 127-qubit quantum processor
I mean, maybe in the theoretical sense, but do conventional computers barely work in practice? Seems a bit of a pedantic argument.

Gil Kalai and others with similar arguments play an important role in the QC community. They keep the rest of us honest and help point out the gaps. But I do think the ground they have to stand on is shrinking, and fast. Ultimately, they might still be right – that much is certain – but it seems to me that the strides being made in error correction, qubit design, qubit control, hardware architecture, and software are now pushing the field into an exponential scaling regime.

To me, the big question is much less whether we'll get there, and much more "what will they be good for?"

bnjemian··on IBM unveils 127-qubit quantum processor
Yes.
bnjemian··on IBM unveils 127-qubit quantum processor
Uncontrolled entropy? If you can repeatedly throw a ball and hit a target but not the bullseye, that doesn't mean you have no control, it just means you need to practice more.

In the spirit of that, there are recent experimental demonstrations of logical qubits:

- NV in diamond qubits (https://arxiv.org/abs/2108.01646)

- Superconducting qubits (https://journals.aps.org/prxquantum/pdf/10.1103/PRXQuantum.2...)

True, a universal logical gate set has yet to be practically realized (at least to my knowledge). But, it seems you're looking for perfection and QC will never be that – even with error correction. Fundamentally, QCs are analog and probabilistic; all discrete states we define theoretically will be approximations in practice if and when they can be demonstrated.

But if the approximation is exponentially close to the target state and that is sufficient for the practical purpose at hand...does it matter?

bnjemian··on IBM unveils 127-qubit quantum processor
Ugh, the point is a fine one but it appears it has to be made:

Validation of experimental theory through the characterization and control of an entire system is not the same as building the same system and simply seeing the final state is what you expect. The latter is much easier and says very little about your understanding.

Here's an analogy: Two people can get drunk, shack up for the night, and 9 months later have created one of the most powerful known computers: A brain. Oops. On the flip, it's unlikely we'll have a full characterization and understanding of the human brain in our lifetimes – but if we ever do, the things we'll be able to do with that understanding will very likely be profound.

bnjemian··on IBM unveils 127-qubit quantum processor
Truth. Not to mention the stunning asymmetry in energy usage. At a minimum, if we can solve an equivalent computational problem using a quantum computer with orders of magnitude less energy than a classical HPC, QC merits consideration. The carbon footprint of data centers is far from negligible.

Computational advantages aren't the only types advantages we should care about.

bnjemian··on IBM unveils 127-qubit quantum processor
Gidney and Ekerå have you covered: https://arxiv.org/abs/1905.09749

Short answer, this probably won't even register as a pitstop on the technical pathway.

bnjemian··on IBM unveils 127-qubit quantum processor
That's not quite it. The power of a QC comes from modeling an exponentially large probabilistic state space using entanglement and superposition. The operations performed by a QC are also different, they can be analog (arbitrary rotations), but circuit depths (i.e. the number of operations) are still expected to be polynomial.

The difference is that the probabilistic state space uses probability amplitudes, which are complex valued and can be positive or negative, allowing for constructive and destructive interference over the probabilities tied to each state. Orchestrate the right kind of interference, and for some problems, you have an algorithm that outputs a solution to that problem with (relatively high probability) in time that, depending on the problem, may be exponentially faster. Examples of those problems include prime factorization/discrete logarithms (Shor's algorithm) and ones in quantum simulation (hence the interest in QC by chemists, physicists, etc.)

bnjemian··on IBM unveils 127-qubit quantum processor
A quanta article (prematurely in my view), tried to dub this Neven's law: https://www.quantamagazine.org/does-nevens-law-describe-quan...
bnjemian··on IBM unveils 127-qubit quantum processor
It depends on your use case. By some measure, the Google collaboration that demonstrated a time crystal was a practical quantum advantage in the space of condensed matter research. But in that space, practical means validating theory with experiment. That's a much lower bar for a quantum advantage than, say, showing that you've implemented a quantum algorithm that can, with high statistical likelihood (e.g. an error rate of 1/10e5), provide fleet routing solutions for a logistics company that are 10% more efficient than any known classical routing algorithm.
bnjemian··on IBM unveils 127-qubit quantum processor
If you added one word of qualification – practical quantum computer or quantum processor – this would actually be a reasonable position to argue.
bnjemian··on IBM unveils 127-qubit quantum processor
I feel like a nuclear war is a much likelier (yet very low probability) event than your computational straw man.
bnjemian··on There may be a steep privacy cost if you park here
Maybe – but I think this assumes that general consumers will forever be unsophisticated with respect to their privacy. That may be the case, but I think it's a bold to claim to make. In my mind, that claim goes against the current trend towards more privacy.

I also highly doubt that a manufacturer of smart products would pass a "we profit off your data" discount down to consumers. Better they pocket the additional profit and then market it as a convenience for the consumer, which I think is precisely what's happening right now.

bnjemian··on There may be a steep privacy cost if you park here
Poison the data!
bnjemian··on There may be a steep privacy cost if you park here
100% agree, I've been saying this for years. I've joked with friends that I'd call it Dumb Inc. It's (mostly) in jest, but I think there's a real mismatch of expectations with a lot of modern appliances with respect to what a consumer needs and what companies use to try and differentiate their products from the competition.

In evolutionary biology, there's this idea that "evolution is stingy" – in most cases, a species doesn't have substantially more capability than it needs to survive (human brains are an often cited exception). I kind of expect that, outside of a computer, most product design should be err on the side of being stingy too. If I buy a toaster, I don't need it to be wi-fi connected so it can send me a text notification when the toast is done. I say this satirically, but I genuinely wouldn't be surprised if one existed. Frackin' toasters.

bnjemian··on Disclosure of three 0-day iOS vulnerabilities
I'd marry mic access, F#$% camera access, and kill bank access. Maybe that's just a personal opinion, though.
bnjemian··on To H-E-Double-Hockey-Sticks with Facebook
I think we should all take a moment to reflect on the true poetry that is that quote.
bnjemian··on I do not agree with Github's use of copyrighted code as training for Copilot
I think a crucial distinction to be made here, and with most 'AI' technologies (and I suspect this isn't news to many people here) is that – yes – they are building abstractions. They are not simply regurgitating. But – no – those abstractions are not identical (and very often not remotely similar) to human abstractions.

That's the very reason why AI technologies can be useful in augmenting human intelligence; they see problems in a different light, can find alternate solutions, and generally just don't think like we do. There are many paths to a correct result and they needn't be isomorphic. Think of how a mathematical theorem may be proved in multiple ways, but the core logical implication of the proof within the larger context is still the same.

← PreviousPage 2 of 3Next →