HNHacker News
TopNewBestAskShowJobs

EthanHeilman

6,075 karma · joined April 8, 2012

Ethan Heilman

CTO and co-founder of bastionzero.com (now part of Cloudflare), author of OpenPubkey (https://eprint.iacr.org/2023/296)

* https://www.ethanHeilman.com

* https://github.com/ethanheilman

* https://hexagon.space/@ethan_heilman

submissionscomments
EthanHeilman··on Bitcoin and quantum computing
That is not what I am hearing from people working on CRQC. A prediction of a CRQC with 10% by 2030 was made by own of the top experts in this field. 2045 used to be the pessimistic outlook by experts with a bunch of experts predicting earlier. Recent work has shown that CRQC is actual 20 times easier to built that previously thought, accelerating all timelines.

We are seeing significant progress in two different types of quantum computers, neutral atom and superconducting qubit.

No one really knows when it will happen, but the chance that it is practically impossible is held only by a small number of experts. Given what we have seen in 2026 has significantly shifted expectations.

EthanHeilman··on Bitcoin and quantum computing
> fork the chain at a snapshot before the attack, patch the protocol, and call that Bitcoin?

It won't work. The only way to authenticate who ones what coins is with signatures. If the signature algorithm is broken, you can't tell who the original owner is to move the coins to a safe signature algorithm.

You need to more to safer signature algorithm before the break, after the break it is game over.

> It’s worth remembering that Ethereum forked for much less

Ethereum could simply return the coins to the original owners. If the signature scheme is insecure, returning the coins just means the attacker can steal them again.

EthanHeilman··on Bitcoin and quantum computing
> 2) to provide objective proof for the true transaction history, anchored in energy expenditure.

Why do you need this if you are willing to trust other people not to steal coins or lie?

> 1) to fairly distribute all coins

Same question as above. If you don't care about perfidy, simply use the honor system for coin distribution.

If you do care about perfidy, then you should probably care about people breaking the law to steal your coins.

EthanHeilman··on Bitcoin and quantum computing
In this case, it seems like we are rolling dice but no one is quiet sure if the dice are fair, how many sides it has and what numbers are written on the dice.

The only thing I am confident in is if it the bigger the fire, the faster the work. I want the Bitcoin community to start the work as early as possible so that it doesn't have to rush because rushing increases the chance of mistakes.

Start early, don't rush.

EthanHeilman··on Bitcoin and quantum computing
"A CRQC is an existential threat to Bitcoin (you might believe this is very low-likehood). Your measurement of this threat should literally be:

(A) How likely you think it is a CRQC appears by a given time, multiplied by (B) How likely it is you think Bitcoin will not successfully upgrade by that time."

It would interesting to survey people about their answers.

My off the cuff answer is:

2030: A=0.05, B=0.01

2035: A=0.50, B=0.001

2045: A=~1.0, B=~0.0

I reserve the right to change my mind on these answers at any point. This is not a serious prediction.

EthanHeilman··on SpaceX files to go public
> See a pattern here? Each failure from a different root cause. So multiple unsolved failure modes, not iteration. It has never reached orbit, never caught a ship, never demonstrated orbital refueling.

Wouldn't it be worse if it was all from the same root cause. That would suggest they can't figure out how to fix the problem. Multiple failures with different root causes provide almost optimal experimental data. I'm not saying the failures are necessarily good for SpaceX, but we are seeing progress and a willingness to push the envelope.

This IPO makes me more worried about the future of SpaceX then experimental rockets blowing up. That said, I can see why some people will invest in SpaceX, SpaceX has a significant lead in the space race and may end up gaining a near monopoly on access to space.

EthanHeilman··on How the AI Bubble Bursts
Not if the bigger models have diminishing returns. Lets say you figure out a way to reduce RAM requirements 100X, but 2x increasing RAM usage by 2x only gets you a 1% increase in effectiveness and 3x does not get you any noticeable increase over 2x at all. Sure you can reduce the price per token, but you might have already saturated the market. Even if you haven't saturated the market, your hardware based moat just got smaller and this is going to reduce your margins even more.

Just noticed that pydry made a similar point: https://news.ycombinator.com/item?id=47574216

EthanHeilman··on Missile defense is NP-complete
It doesn't have to be, defender reveals everything and attacker chooses best strategy.

1. The defender could use both electronic and physical decoys, use air and sea mobile platforms that are always in motion and are difficult to track.

2. The defender can fire at decoys, to convince the attacker the decoys work when they don't.

3. The defender could mix in cheap decoy interceptor missiles that miss so the attacker concludes defenders need 10 missiles to intercept when the real number if 3 and the attacker thinks the defenders are running low on interceptors, when in fact the defenders have held most of their interceptors in reserve.

4. Defender can pretend that expensive systems have been destroyed so that attacker adapts their strategy. For instance, if your defense hinges on a small number of extremely expensive fixed X-band radars and the attacker targets them. Allow some of them to be appear to be destroyed when in fact, you have disassembled them and moved them somewhere else to use later in the war.

I see no evidence anyone is doing any of this today, I'm not making any sort of claims about deception operations in the current conflict.

EthanHeilman··on Nobody gets promoted for simplicity
Seriously? I worked at startups and research institutions. We trained people on interviews. I know Google used to invest quite a bit on interview training.
EthanHeilman··on Nobody gets promoted for simplicity
Complete agreement. "Excellent answer, that is what I would do as well, now what if we wanted to build it in-house?"
EthanHeilman··on GPT‑5.3 Instant
How likely is that they dropped this now to push the news story about quitGPT out of the headlines?
EthanHeilman··on Tesla is committing automotive suicide
If Tesla has lost the advantage in battery tech, that is unfortunate and speaks poorly to Tesla's long term strategy. Reclaiming this lead would be an important strategic goal and I disagree with that not being prioritized.

> Why? Other robotics companies have been doing it for longer. Is Optimus better than Atlas:

Atlas costs about half a million dollars, targeting a price tag of $160,000 once mass produced, and assumes the user will be able to do some maintenance.

Optimus is targeting a price tag of $30,000, but probably costs around $80,000 to produce. It is plastic, it is cheap, it doesn't work.

Atlas is better than Optimus but all measures. The advantage of Optimus so far has been, the mass production-->usage until failure-->improvement cycles that are already underway. Tesla is, as an extremely high cost, slipping on every single banana peel first and this is clearing a path for other companies to learn what doesn't work when you switch from functional over-engineered robot to barely functional robots that can be mass produced.

Telsa isn't alone in this space, but they investing a lot and trying to cut corners. So much of engineering is learning the corners you can cut and the corners that cause a battery fire after 8 weeks of use.

EthanHeilman··on Tesla is committing automotive suicide
> Tesla has 4 different, 4 person cars. It's redundant.

You are spot on, it makes sense to have the Model 3 (economy sedan) and Model y (upmarket crossover SUV).

My question here is why did Tesla have four 4-person cars in the first place? If you wanted to streamline engineering and supply-chain why have Cybercabs instead of using the model 3 or model y as the base? Why split the company between Optimus and making cars?

Cybertrunk does make sense, it is a technology demonstrator and test article filled with all the new ideas and tech they are going to build into the next generation. They get data on people using it by selling it to them.

What you say is a sound strategy for Telsa to peruse, but they don't seem to be perusing it.

EthanHeilman··on Tesla is committing automotive suicide
EVs are a solved problem, but as amelius notes the real tech is the battery. Tesla + Panasonic has a built in advantage in terms of battery manufacturing. Tesla has a massive amount of capital, if they put it into reducing and scaling manufacturing of vehicles and batteries, I think they could probably win. Now maybe Telsa has looked at the numbers and decided they can't win and are choosing to pivot rather than die a slow death.

I don't think that is what is happening here. Instead, Tesla is continuing the strategy that brought them to this disaster of going all in on driverless. That isn't a bad strategy, but if they get the timing wrong a third time, they destroy the company and they have gotten the timing wrong on this twice already. This strategy has two downsides:

1. AI has no real moat and Tesla has largely pursued commodity sensors, meaning that other than EVs+battery tech (which Tesla appears abandoning), robotaxis have no hardware or software moat.

2. They could use network effects to win, in which case their competitors are not other car companies but Uber and Lyft. Uber has been pursuing the same long term strategy at Tesla.

Now by itself, going all in robotaxi, is risky but could work if they time it right. Tesla isn't going all in on robotaxi since they are splitting the effort between robotaxi and Optimus robots.

It is likely that the experience Tesla gets with Optimus robots will help other robotics companies, but unlike robotaxis where the timing might (but probably won't work), the timing is clearly isn't right for Optimus.

It seems like the motivation here is that Musk is aligning Tesla to a narrative that justify the absurd stock price, even if that narrative isn't reality.

EthanHeilman··on Iran Goes Into IPv6 Blackout
Thanks for your comment, I know the ionosphere is used in Electronic Warfare but I didn't realize it was so limited in frequency.

Is there really is no way to reflect signals off the ionosphere out of phase so after reflecting they interfere into a higher frequency?

EthanHeilman··on Iran Goes Into IPv6 Blackout
Yes, but it is more difficult than jamming a typical radio antenna because the starlink uses a directed beam rather than a omnidirectional radio broadcast. This either requires enormous amounts of power, targeting the satellite itself with a directed radio beam, or getting between the satellite and the ground station by bouncing a signal off the ionosphere.

The above is for jamming directed beams in general. It is likely that starlink has a number of other jamming countermeasures.

EthanHeilman··on Opus 4.5 is not the normal AI agent experience that I have had thus far
I haven't seen an AI successfully write a full feature to an existing codebase without substantial help, I don't think we are there yet.

> The only question is how long it takes to get there.

This is the question and I would temper expectations with the fact that we are likely to hit diminishing returns from real gains in intelligence as task difficulty increases. Real world tasks probably fit into a complexity hierarchy similar to computational complexity. One of the reasons that the AI predictions made in the 1950s for the 1960s did not come to be was because we assumed problem difficulty scaled linearly. Double the computing speed, get twice as good at chess or get twice as good at planning an economy. P, NP separation planed these predictions. It is likely that current predictions will run into similar separations.

It is probably the case that if you made a human 10x as smart they would only be 1.25x more productive at software engineering. The reason we have 10x engineers is less about raw intelligence, they are not 10x more intelligent, rather they have more knowledge and wisdom.

EthanHeilman··on Opus 4.5 is not the normal AI agent experience that I have had thus far
Even if you are going green field, you need to build it the way it is likely to be used based a having a deep familiarity with what that customer's problems are and how their current workflow is done. As much as we imagine everything is on the internet, a bunch of this stuff is not documented anywhere. An LLM could ask the customer requirement questions but that familiarity is often needed to know the right questions to ask. It is hard to bootstrap.

Even if it could build the perfect greenfield app, as it updates the app it is needs to consider backwards compatibility and breaking changes. LLMs seem very far as growing apps. I think this is because LLMs are trained on the final outcome of the engineering process, but not on the incremental sub-commit work of first getting a faked out outline of the code running and then slowly building up that code until you have something that works.

This isn't to say that LLMs or other AI approaches couldn't replace software engineering some day, but they clear aren't good enough yet and the training sets they have currently have access to are unlikely to provide the needed examples.

EthanHeilman··on Scientists edited genes in a living person and saved his life
Science works better with collaboration. If China and Europe have to take up the slack from a self-inflicted wound to US research, science as a whole suffers. There is far more science to do than scientists and when science funding gets canceled the number of scientist decrease.

After the US canceled the SSC, Europe built the LHC. The LHC isn't a one for one replacement of the SSC, but it would probably not have been built if the SSC had been built. What scientific projects would Europe have built instead? What did we lose?

EthanHeilman··on Scientists edited genes in a living person and saved his life
With Trump it appears a window is closing on the development of important technologies and research. I doubt we will enter a new dark age, but in some areas, progress is likely to slow, which in turn will be used as evidence that it isn't worth funding, which will cause funding cuts and result in even slower progress. Everyone is racing to get stuff done, because there might not be the money tomorrow. It seems likely to me that longer term, more ambitious projects are probably being sidelined because there isn't time to complete them before the pencil pushers in the whitehouse defund them.

Look at how the Nixon administration and congress gutted NASA. It took nearly 40 years to crawl out of the hole in the roadmaps that their shortsighted stupidity created. We could have had reusable rockets, aerospike engines and Nuclear Thermal Propulsion in the early 1980s. Instead we got halfway measures like the space shuttle and the ISS that both ate budget but didn't create the required innovation for lower cost to orbit.

EthanHeilman··on LD_PRELOAD, The Invisible Key Theft
LD_PRELOAD is so useful for non-malicious stuff that I hope it doesn't get a reputation as a bad thing to find on your system. That being said, I agree with you and also disagree.

From a defenders perspective, you have lost if an attacker has root access on your system. You are right. Consider instead the attackers perspective.

To an attacker compromising and system and gaining root is just the first step of a many step process. One of the hardest steps is modifying the system to silently collect and exfil secrets and data that is valuable to you. Let's say you want encryption keys and only keys, how do you get them? For the sake of example say they are stored on the file system and you want to exfil them as they rotated weekly. Do you write a program with a cron job that checks once per day and uploads them? What if three months later they switch from rotating their keys once a week to once every two hours?

1. How long does it take you to notice your missing most of the keys and what is the cost of this failure?

2. Once you notice you aren't getting all the keys, you need to figure out why. This can take time and money. Do you access the compromised machines again? What if you can't get back into the machine again to figure what happened?

3. Once you figure out why, you need to deploy a patch to your exfil kit. This again costs time and money. What if you didn't test it properly and it breaks the compromised host and exposes your entire operation? You might have to push this one to thousands of compromised machines.

Instead, use LD_PRELOAD to hook filesystem writes, pattern match the key format on and exfil the keys as they are written. Since the hook is environment variable based, it can survive changes to the targeted program. Granted there are other approaches as well, but LD_PRELOAD is simple, powerful, flexible and often used for non-malicious things so it doesn't immediately trigger alarm bells.

EthanHeilman··on Faith in God-like large language models is waning
> The slowing pace of improvement at the bleeding edge of generative AI is one sign that LLMs are not living up to their hype. Arguably a more important indication is the rise of smaller, nimbler alternatives, which are finding favour in the corporate world.

This seems to be saying that SLM (Small Language Models) are a sign that improvement in LLM (Large Language Models) is slowing. I don't know that improvements in LLM are slowing, I really mean it, what is the data one way or the other. I've been trying to investigate this question and I don't have a clear answer. The rise of SLM doesn't seem like evidence either way.

EthanHeilman··on OpenAI eats jobs, then offers to help you find a new one at Walmart
In my personal experience I've seen:

- OCR eat a good chunk of data entry jobs,

- Automated translation eat a number of translation jobs,

- LLM have eaten quite a few tier I support roles.

I don't have numbers tho, maybe people are still doing data entry or hiring translators on mechanical turk.

EthanHeilman··on God created the real numbers
How would you define provable without first assuming a set of axioms?
EthanHeilman··on God created the real numbers
As far as I can tell this is not a misconception, I am just not communicating clearly enough. I figured I could make the point at a higher level.

You can construct a statement in an axiomatic system complex enough to express arithmetic:

1. Which is true,

2. AND which can be proved to be true under the assumption the system is consistent,

3. and which can not be proven true within that axiomatic system because you can't prove consistency of the system itself within the axiomatic system itself.

The crass positivist position would be that such undecidable statements are meaningless, i.e. neither true nor false. The incompleteness theorems provide a strong counter-example: a statement which can not be proved within that axiomatic system, but is true (if the system is consistent). The positivists must then amend their position to include things which can't be proved within the axiomatic system but can be proved true. This resulted in the watered down late-positivism.

EthanHeilman··on God created the real numbers
> What Godel actually proved is that given a consistent formal system capable of representing arithmetic, there are statements in the language of that system such that neither the statement nor its negation can be proved in the system.

It goes what step further, there are statements that are true but can not be proved.

> Godel's theorems are notoriously liable to misinterpretation, and to be honest this sounds like one.

What do you think was Godel's philosophical motivation for investigating incompleteness?

"Many philosophers and logicians have explored the significance of these theorems. However, they would often find a reason to distance themselves from the interpretation their author, Kurt Gödel, ascribed them. Gödel believed that his results provide a strong argument for the objective existence of a rationally organized world of concepts, which can be to some extent described by a deductive system but cannot be changed or manipulated (cf. [12], p. 320)." [0]

You can argue that Godel misinterpreted incompleteness and There is a case to be made for that. Godel probably did not misinterpret his own philosophy about incompleteness.

> Godel's proof itself can be formalized, so I don't see how it places a limitation on formalism in general

It uses formalism to show that formalism is limited and that the positivist notion of truth does not include things which are true but can not shown to be true (within a axiomatic system capable of expressing arithmetic).

[0]: Kurt Gödel and the Logic of Concepts (2024) https://arxiv.org/html/2406.05442v2#bib.bib12

EthanHeilman··on Why haven't quantum computers factored 21 yet?
any sources? I'd be interested to read a critique of quantum error correction
EthanHeilman··on Why haven't quantum computers factored 21 yet?
You brought up gate noise, there has been quite a bit of progress on that problem.

> so we should totally be able to factor 21 (or larger numbers)…. When?

Just because we solve one problem doesn't imply all the problems in QC are also instantly solved. I guess it does if you assume noise is the only problem and once is it solved the engineering is trivial. That is not the case. Even assuming all foundational problems have been solved, figuring out how actually engineer and also mass produce large numbers of gates, will take a while.

As the article pointed out, going from 15 to 21 requires a 100x increase in gates.

As the article that you posted under says:

"Because of the large cost of quantum factoring numbers (that aren’t 15), factoring isn’t yet a good benchmark for tracking the progress of quantum computers. If you want to stay abreast of progress in quantum computing, you should be paying attention to the arrival quantum error correction (such as surface codes getting more reliable as their size is increased) and to architectures solving core scaling challenges (such as lost neutral atoms being continuously replaced)."

EthanHeilman··on God created the real numbers
> The fact that something has been discussed for thousands of years also has nothing to do with whether there are good reasons for believing it, e.g. the Earth being flat or stationary at the center of the universe. People can be wrong for thousands of years

I agree people can be wrong for thousands of years, but a flat Earth was conclusively disproved early on. I'm not talking about random people debating this issue, I'm talking a continuous intellectual process over thousands of years to understand the nature of mathematics.

It may turn out that various theories put forward will be shown to be wrong. This has already happened, Godel's incompleteness was a major blow to people that argued that mathematics did not have an independent reality because it was just a logical game defined by axiomatic rules. Godel showed that axiomatic systems beyond a certain level of complexity have a reality beyond that logic can investigate.

EthanHeilman··on Why haven't quantum computers factored 21 yet?
Quantum error correction addresses this problem and we now have tech demos showing that we can build scalable QCs with surface codes [0].

[0] https://scottaaronson.blog/?p=8525

← PreviousPage 2 of 34Next →