Now, in web-facing systems, the biggest concern is the low-level crypto libraries, which need to be written in assembly to avoid timing attacks. These (and related libraries) have seen multiple security advisories in the past year, some related to unsafe (I think?) and some related to logic errors. But overall, those advisories have mostly been lower priority, or affected mostly niche use cases.
If you only have a small amount of unsafe code, then it's possible to audit it more aggressively. And there's only a limited amount of space for bugs. So what I would expect for Rust is a trickle of security bugs, many of them niche or low severity, that slows down with time.
[1]: https://projectzero.google/2014/08/the-poisoned-nul-byte-201...
If any single byte overflow can be converted into an RCE with minimal human effort then we have no choice but to prove the whole software stack is memory safe.
I wrote https://github.com/anacrolix/possum for this exact use case. It has a few rough edges but it absolutely hauls ass for caching on disk. Incidentally it uses sqlite for transaction/TTL metadata which is more suited to sqlite but it's still the bottleneck.
sqlite transaction commit overhead is too high and uncompetitive without batching but it's a design flaw specific to being multiprocess.
But she didn't think it was relevant that in the book the martial island then easily conquers the pacifist one and destroys their way of life. I did. To me it was a lesson that "si vis pacem, para bellum".
I see her position in the demands of these tech workers. In the fullness of time it wasn't ethical for the peaceful islanders to refuse responsibility for their own defense, as it set the stage for the downfall of the society they loved, in the name of being caring people. A modern phrase for this is suicidal empathy. Short sighted works too.
If you think you live on the martial island, the ethical move is to stop helping to build their whole economy, not just their military.
It does not mean “multi-tasking machine”, it means: “person who can’t convince themselves to start, or stop, any task with the exception of emergencies or things they might believe are novel”.
I’ve never been able to control what I work on, no discipline or desire has helped.
The idea of trying to do 10 things at once makes me physically ill.
I've started exploring the use of AI and there is simply no sense of joy or achievement, even if I actually can do much more with AI now than I did solo in the past. Back then, every little thing I managed to figure out myself (either by reading the documentation or by experimentation), every little win brought me a bit of joy and made me want to celebrate. All of that is gone, and the current job market simply does not allow me to give a shit - every job post demands AI, and a lot of them will want you to run multiple sessions in parallel. @MiaAI_lab on X said that "no ADHD is a disadvantage in AI", and I tend to agree. I interviewed for a job a few days ago and the girl on the other side is not even senior - she tells me she has ADHD and is able to drive 10 agent sessions in parallel. To me, it sounds like ADHD means she's not going to give a shit for any of the running agent sessions, which is somehow a superpower since 2025. I'm in my 40s with more than a decade of experience and I don't think I can ever work like she does. It's just really scary and disheartening because I feel like I have become unemployable.
"It is the capacity for maintenance which is the best test for the vigor and stamina of a society. Any society can be galvanized for a while to build something, but the will and the skill to keep things in good repair day in, day out are fairly rare. […] I read somewhere that in ancient Rome a man was disqualified as a candidate for office because his garden showed neglect."
Absolutely.
However, we've also never poured more in, to such centralised infrastructure.
I've actually never worked in any company that has had a communications outage for 36 hours in my life; and I've worked in places we had a dancer IRCd running on an old laptop under the faux floor in the server room with Solaris as the main communications tool... thing is: it was so simple, when it went down we just fixed it.
Technology companies have very successfully convinced us that we should outsource responsibility, because there are economies of scale. Then they ask for a free pass because "scalability is hard".
You don't get to say scalability is hard when you've been pushing and propagandising centralisation for over a decade, while having thousands of staff on million-dollar yearly TC packages.
Everyone here will say "nobody got fired for buying x- because everyone does" (Slack, IBM, Microsoft, AWS), and you can use it as an excuse... but you know what happens when we over-rely on something like that and it fails?
Economic crashes. That's exactly how the 2008 financial crisis happened, everything depending on something that had "absolved" responsibility. Such a thing has ruined the prospects of an entire generation of people, and I'm actually becoming really upset that we think we're being clever bastards because "oh hum, well, I can't be held responsible for total systematic collapse, everyone else was doing it!"
The thing is that people want a nice life, and most smart people know that martyring themselves will neither save the world nor give them a good life.
The rational thing to do is walk the rim of the proverbial canyon, staying close enough to make good money but far enough away from the Big Questions to never feel like they should become your responsibility to sort out. Because once you agree it should, it’s a self respect hit to walk away without doing something, or do something at personal cost and probably fail to achieve your goal anyway.
Especially since we are not the best people to fix social problems, which very many things are.
This resonates with me. It feels insulting when one is told not to use your own brain and judgement, and to defer entirely to machines. It puts humans below AI. It's dehumanizing.
Don't force business to care about something that they don't (it won't end well for you) Equally Don't force yourself to care about something that you don't (it won't end well for you)
If there's an overlap, find a way to prove your value in relation to that area and use it as proof of your value and leverage that to be able to deliver more of what you value.
See guy putting out fire getting praise vs one that wrote their stuff correct the first place and it didn't caught on fire.
Wasn't even argumentative, just quietly gave a shit.
There are plenty of smart people, but you probably aren't paying them enough to argue and play politics with the imposters that you have also hired. If doing something well also requires convincing a (product) manager that it was their idea, then no one will be doing anything well. The best are going to checkout.
If you get rid of the imposters and lessen the cognitive burden of playing politics to get anything done, it will appear as though all your smart folks started caring.
A CEO might know that keeping smart people is a good thing. But the incentive structure for the directors and other middle management just make it impossible because smart people get expensive and tossing them looks good for those middle management bonuses.
Smart people are opinionated, and management thinks very highly of themselves so they can't have anyone with "bad vibes" around.
Of course there are more. So what can you do?
Yet the thing is rate at the top of how AI have revealed to me that in an organization of 400 people, 3 give a shit. I literally can’t find anyone who gives a shit in insanity large organizations. Not low paid entry workers, not partner or VP, not CEOs. All just a bunch of people who probably never gave a shit to begin with, but AI put that at the forefront.
In a perfect world, the second and third group would be the same, but since impressions of competence are lossy as information travels up the hierarchy, optimizing for the impression of competence (i.e. “visibility”) will always be more rewarded than actual competence.
IMO, the issue is that AI is reducing the cost of building and modifying systems much faster than it's reducing the cost of understanding them. For example, an engineer can work on a problem or project for a week that could otherwise be done by AI in an hour or so. But, that week spent was not "wasted time"; it was understanding what it is they are building, discovering undocumented assumptions, knowing about different dependencies, developing intuitions about the system and how it works, etc.
This causes comprehension debt. Yes, the system does ultimately work, the tests also pass. But the level of comprehension goes down significantly.
That said, AI can also help reduce this debt by asking it what you need help understanding.
A good blog post on this: https://larsfaye.com/articles/ai-coding-will-prevent-experti...
They are sidelined, because giving a shit is not reflected in OKRs and giving a shit often equals standing in a way of "good enough" as judged by a middle manager and thus committing mortal sin of not focusing on getting to market faster.
It was like in the 2000s, blogs wanted to share informatin but also be funny, something similar to a talk in a event? Later they only focused on SEO, so they optimized every word and paragraph, removing jokes, etc. I remember noticing this with the Digital Ocean docs, that ranked 1 for every tutorial in how to install stuff on servers.
Anyway, good times! I miss them.
Qwen 27B really punches above its weight. 27B is small enough to fit on a Mac Mini / Macbook Pro which is very exciting.
Starting to see more promising examples of companies post-training. From my evals [1], I've tested 3 so far:
1. Fireworks' Ember-1 (postrained on Kimi K3, improved score by 10% in my evals)
2. Particle AI's Particle-1 (fine-tuned on DeepSeek V4.1 Flash, also a 10% improvement and a bit cheaper and faster)
3. Nex AGI's Nex-N2.6 (post-trained on GLM 5.3 Flash. Didn't show an improvement in my eval, and had a couple runaway loops that made it expensive).
With Thinking Machines' western open model release (Inkling Small) a few weeks ago, I expect we'll see more of this, especially out of western companies.
[1] https://plotly.com/blog/nex-n2-6-plotly-data-analytics-bench...
Sam Altman, would you actually believe anything he said if he had an incentive to lie to you?
Rieman hypothesis story was not secret. How do you even entertain this stuff for a second? Even if it were in your financial interest to not understand it’s just so very stark.
The AI companies will tell whatever lie helps them and won’t blink when caught and there will be no consequences because we’ve seen it. Repeatedly. There isn’t a grey area…
So if I am getting this right?
+-------------------------------+--------------+--------+--------+----------+-------+
| Lab | XL | L | M+ | M | S/XS |
+-------------------------------+--------------+--------+--------+----------+-------+
| Google (Gemini 4 Family #1) | | Argon | | | |
| Google (Gemini 4 Family #2) | | Carbon | Neon | Hydrogen | |
| Anthropic (Claude 5.2 Family) | Fable/Mythos | | | | |
| Anthropic (Claude 5.5 Family) | | Opus | | Sonnet | Haiku |
| OpenAI (GPT 6.2 Family) | Astra | Sol | | | |
| OpenAI (GPT 6.1 Family) | | | | Terra | |
| OpenAI (GPT 6 Family) | | | | | Luna |
+-------------------------------+--------------+--------+--------+----------+-------+
Notes: * Gemini 4 Family
- Argon -- ++Train --> Carbon
- Carbon -- Distill --> Neon, Hydrogen
* Gemini 3 Family
- 3 Pro -- ++Train --> 3.1 Pro
- 3 Pro -- Distill --> 3 Flash
- 3 Flash -- ++Train --> 3.5 Flash ... -- ++Train --> ... 3.8 Flash
* "S/XS" classification is due to weirdly low pricing for models by Ant/OAI and not by raw perfHaiku < Sonnet < Opus < Mythos/fable
I think openai has a similar system with astral bodies:
Luna < Terra < Sol < Astra
Camry = Kammuri = Japanese for "Crown"
Crown = English for "Crown"
Anthropic spent a lot more time editing their pre-training data with higher-quality material. For example, they used actual books for pre-training, because books are generally run through human editors for higher quality. (this is from the copyright lawsuits)
[ ... ]
Meanwhile its obvious that companies like Google would use trash data like email and chat logs or Internet forums for pre-training. Just complete trash data.
Google has been scanning every book they can get their hands on for 20 years or more. They have a bunch of patents on non-destructive book scanning gear. Surely they are using that data to train their models.Because probably there is no system. They just train a model, and find out that it's "kinda good", and that's it, they publish it. The models do not constitute even a partially ordered set, nobody can predict their ranking for your particular tasks.
To the point that I now use a workflow of delegating from Opus or Fable to Gemini4 to write commit messages and PR descriptions, it mostly solves the Claudese problem.
I do apologize in advance for putting you in spot like this, but there are two things I have been quite curious about.
1. Since 60% of the benchmarks contributing to AA are now private only, how do you guys ensure transparency? More precisely, how can someone verify the transparency in scoring?
2. There are also disclosures about custom benchmarking services offered to AI labs, what does AA do to keep the custom benchmarking services not getting access to the same/similar data that is used in the private benchmarks?
I think internally, Google, Anthropic and OpenAI are thinking "It's so easy, you just use Neon for X, Carbon for Y" - but most of the world is trying to map Sol, Astra, Neon, Opus, Fable, Mythos, Carbon, Haiku, etc.
I just hope they finally land on a naming scheme they are happy with and just stick with it. But I have my doubts. Perhaps they are all benefiting from this confusion.
Other companies would use completely raw internet data for pre-training, which is basically trash. Anthropic's secret sauce is that they filtered and edited down their data for only the highest-quality material. You really don't need quadrillions of tokens to pre-train, as that's just random noise. A smaller-but-higher-quality data set is better in every way.
Meanwhile its obvious that companies like Google would use trash data like email and chat logs or Internet forums for pre-training. Just complete trash data.
This is a classic case of "less is more" and "garbage-in-equals-garbage-out"
With the frontier getting more jagged, we might see more of this 'less intelligent overall but useful for the frontier tasks'-type models, meaning picking a model will only get more difficult from here.
(FD: I work at AA)
No wonder they went for the 5.5 naming, instead of the .1 increase. From just the benchmarks, it seems like neon is basically on par with Opus 5.5 which came out 2 months ago.
Is it because Anthropic put a ton more effort into the character of the model that it basically roleplays as a human and that why it tends to think better when giving tasks that it might not be trained heavily on.
I have used Carbon more than most model to try and get the same results but it just feels like i'm talking to a smart AI. Which is fine if you are asking questions, researching an idea, but if you actually want to work on something, build something, it just falls apart. Like Opus can run for days on a task and it feels like it understand what it was doing 2 days ago even when it might not be in it context while other models feel like they breakdown and get lost or falter as the task grows.
I'm not sure it's much better than 4.1 Flash vs 4 Pro -> if 4.1 Flash is better temporarily when Pro is typically significantly better, people are going to be confused.
I'm not really sure how you get around this problem easily.
It's not going to age well when the next Flash-lite model launches, but maybe they will call that `Gemini 4 <Element>-lite` - who knows...
This is probably the set of benchmarks most people care about: https://www.vals.ai/comparison?m=claude-fable-5-2,cobalt-c-h...
Overall, performs just under Fable (above Astra or Sol 6.2 or Carbon - slightly ahead of Opus 5.5).
Latency is by far the best for the level of quality.
It's about 2x more expensive than Sol 6.2, but significantly better and faster. You get what you pay for.
Started off with me searching for what Chinese OEM made the product. Then tried out existing GitHub projects to see if it works but they were failing.
Spent 15 minutes to write a script that started their desktop software on a different network namespace and ran tcpdump in it. Captured exactly only the things i needed without any noise. Turned out it was using a different server url and credentials than the GitHub projects. Opened Ghidra to inspect the dlls that matched the strings i was looking for.
After that, I got lazy and handed over the task to Gemini to find the username and key they use to authenticate with the servers. It took about 30 minutes or an hour. A lot of it was me changing models because the llm servers were busy. It simply wrote a little script that ran through various strings it fetched from the binary to see if it can generate the authentication credentials.
Now my Home Assistant will be detecting if my car is parked in the parking lot or not. Life is good.
Companies don’t play by the rules anymore, neither should we.
Not too long ago, I successfully reverse engineered an abandoned audio effects plugin, as well as a discontinued instrument with an onboard DSP for virtual instrument modeling. I had no idea what I was doing, but it was still amazing to get to open the black box and see what's inside and how it worked.
The plugin is nearly 1:1 as far as sound, and while I'm not going to use the code or the techniques of the instrument, peeking into them gave me the insight I needed to start trying to replicate the functionality on a modern DSP.
I have my issues with generative AI as a technology, but busting open propietary software - particularly when it's abandoned and unsupported, is one of the cooler use cases I've seen for it.
Some people have cleverly parlayed this new fad into money, but this is typically a result of their sales techniques rather than their software.
Runs great on my linux box
Windows-1252 encoding has a representation for lowercase German sharp 's' (ß), but not one for the uppercase. If you upcase a string containing a sharp 's', it gets normalized to "SS", which increases the length of the string by one character. The 1252 case conversion code doesn't handle this; it looks like it checks for headroom in the buffer one character at a time. So if the last character of a string you upcase is ß, it looks like you get a one-byte overflow (of 'S').
If that's really the bug (there's a claimed RCE POC), the heap grooming for it must be really fussy! I couldn't find any other character pairs that expand like this.
You can feed strings in 1252 to a Rails program through multipart POSTs, which get to pick their encodings.
The great thing about this bug, apart from how it illustrates how doomed we all are (if I'm right), is that it should by rights have the best vulnerability name of the year, because you get to work 'ß' into it.
Yes
The main reason I see to not use Envoy is that it's very annoying to set up and build, and requires familiarity with things like grpc/proto/low level networking to get right. But if you're using grpc and http2 I think you're generally past that "infra overhead" threshold already and it's going to better serve your needs.
I have pretty much only worked at big tech or my own infrastructure startup, and did some systems/infra work at Google using Envoy for a while. I can't say it's 100% better than nginx but I can say that I've evaluated nginx and a lot of software like it (ie with a similar customer base, ecosystem, and history; like coturn) over the years for my personal use, and every time I've immediately been annoyed by constant security issues or poor http2/L4 networking support that I don't see when I use what Netflix/Google/etc use.
The comment you replied to is shortcutting the joke, assuming the common riff of becoming a carpenter is well known, and jumping ahead to asking about tutorials for woodworking as part of that process.
But when newer software has the same vulnerability, you do not see the same kind of comments.
There always seems to be a latent stance in IT that newer is better.
New code means new bugs, and losing the many years of accumulated bug fixes and hardening present in many critical pieces of our infra by rewriting in rust is not a desirable tradeoff. The new world we live in where fv has become many orders of magnitude cheaper also makes rust seem much less attractive.
the certificate for it was issued on the 2nd of November
if you were scanning CT logs you would have had a 16 day head start on RE'ing this
there are more examples in the current CT logs
https://api.certspotter.com/v1/issuances?domain=nginx2root.c...
If I wanted to build my own software from source every time I'd be running Solaris like a virile man from 2004.
https://xint.io/blog/nginx2root-remote-code-execution#23-com...
Workarounds if you're vulnerable and can't update immediately:
https://xint.io/blog/nginx2root-remote-code-execution#33-eme...
There are two internets. One of them is open, less-commercial, and largely built and maintained by people whose first job is not the work they do to maintain this internet. That group includes a ton of open source software, millions of small servers scattered around the world (virtually private or otherwise), and is generally best-effort.
The other internet is fully commercialized, often ad-driven, fully integral to the data broker economy, and centralized behind a few companies that pay as few people as possible to satisfy the occasional PR statement that, "we take security seriously and...".
The next stable equilibrium here is not "we get a secure and open internet". The next stable equilibrium here is, "we have no open internet, and the internet is now managed by just a few companies."
We're already halfway there. Google and Microsoft combined already account for somewhere around 30% to 40% of email traffic. 30% of all websites are now behind Cloudflare. A few companies dominate social media traffic. Amazon and Shopify combined are half of US e-commerce.
Every single short-response-window exploit like this just pushes a few more percentage points of traffic into a handful of companies. Cheer for that, I guess.
When I have time to daydream, it's increasingly about building a new, underground internet, and how best to disadvantage bots and BigCo conglomeration from the start.
Worse patching adds instability into a system, especially when you don't have much time to test.
Also in the past attackers were willing to sit on exploits longer and not do things with them, now they have to act fast before someone else finds it, leading to more events faster wearing your security and patching teams out, and making users life suck.
Things are going to get much worse before they get better.
Given that in this case the workaround is as simple as disabling http2, there is honestly no reason why F5/Nginx or any of the reporting teams couldn't at least put out a PSA to disable http2 for the time being before basically telling the whole wide world this exists and how to exploit it.
Especially for software like Nginx which powers a huge part of the internet.
Are we going to discover in a year that every server online this month had a rootkit installed? There’s got to be thousands who are too small time to wipe and reinstall.
If anyone gets a hint of the space a vulnerability occupies, just point your agent at it and you have a pretty good chance of finding something. (See java4shell and the NYT reporter who was able to replicate the problem with $200 of tokens)
Is this some kind of record for bug collision?
"The Box" points out that shipping did not convert to containers as much as replace ports and shipping lines with new builds built for containers. The Port of San Francisco used to be a major port. But, when containers came along, a new container port was built in Oakland. One that didn't have an entrenched longshoreman's union. A few things helped. BART was under construction, and digging up dirt that had to go somewhere. That dirt, as fill, became the container port.
London went in a similar direction. London never got a container port. Instead, a minor non-union port on the eastern coast was built up into a container port. The London Docklands died, to come back decades later as office and residential space.
Containers needed all new ships, and made existing ships obsolete. Some shipping lines converted; others were superseded. There was no one big winner in containers.
It took quite a while to get trucking and railroads on board. Railroads had tried various intermodal schemes - trailer on flat car, trailers that could run on rails, and such. These took less infrastructure for loading and unloading than the big maritime boxes. But eventually the dumb boxes won out. Union Pacific even increased its clearances to allow double-stack container trains.
The US longshoremens' union got deals for lifetime support for their existing members. But there were not many new people entering the industry. If you visit a container port today (if you can get in) there won't be many people around. Here's what one looks like.[1] That's not a render; it's the port of Rotterdam, a decade ago. (The video is sped up; the AGVs are not that fast. This is real speed.[2]) All recent ports are at least that automated.
I guess the result for AI is don't try to get rich inventing AI. Try to get rich using AI to enable things that weren't previously possible. I think that space has barely been explored - so far most of the activity is not net new things, but just trying to do the same thing we always did before, but faster/cheaper.
It's well argued and the arguments are worth considering.
It feels like all the worlds most powerful companies are pushing these things on us to make us stupid and reliant. This timeline sucks so much.
We actually did significant research before llms. You cannot convince me that research and progress is getting done now, outside of like 10 companies.
Every is just using llms to cheat at their jobs and avoid tackling hard problems. I fully expect there to be a complete plateau in human progress because of these things, while its dressed up as the opposite.
Everybody is cheating on their jobs and while thinking to themselves "I bet theres some mega smart people out there using these llms to cure disease" while those super smart people are also just being lazy, cheating and letting their brains atrophy and also hoping someone else is taking the hard stuff seriously.
Gg idiocracy is coming.
Yes Anthropic will be leading for a while with the best model; but first the Chinese models, yesterday's OpenAI's Sol & Astra, and now even Microsoft getting Opus 5 performance. Remember that even Anthropic's own Opus 5.5 is _way_ more efficient than the models that were active in their “profitable” quarters.
I do not see a business that is going to do triple digits growth consistently to justify their valuation.
It is just too much of a fungible utility — and there is too much demand for any single approach to capture a market. Hell, AI is the perfect tool for any customer to 'uncapture' themselves.
- A big contributor to Postgres OSS work
- A big contributor to Python
- Ships VS Code and Monaco, on top of which many, many startups have built their products and VS Code (or offshoot) is used by every startup I've been in and increasingly seeing adoption in enterprise
- Ships TypeScript, a cornerstone of the modern JS stack
- Owner/operator of GitHub which, despite its issues that we all know of (no point trashing them here), still holds a massive network effect with the majority of open source projects hosted on GH, running billions of Actions minutes/mo. I'd guess.
Microsoft and Oracle are very, very different.
Microsoft doesn’t need any of that. They have the market share, distribution, brand awareness, and a lot of their potential customers already use Microsoft products somewhere.
The phone comparison doesn’t apply, IMO. Phones are entire ecosystems and they’re very sticky. People don’t switch phone OSes casually because they only buy a phone every several years. Getting locked into a phone you don’t like is too risky.
AI plans are not like that. Not yet, anyway. Vendors push for contracts and one year lock ins via discounts, but individuals and nimble companies are changing AI providers all the time.
Even my wife’s big company technically has a contract with one AI provider for the company, but teams across the company are getting extra spending to use on Claude or other models of their choice and the favorites are changing every few months.
Trying an AI model is low risk. You can do it for a month or even just earmark a token budget and try it. Nothing like buying a phone that you’re stuck with for years.
In this case there is huge demand from companies who consume API tokens at currently high prices who do not want to buy from the labs and who find that existing open source models are starting to lag the frontier too much, or are not actually that cost effective once accounting for the amount of tokens they consume.
The audience here is enterprise looking to save money, and there is no natural monopoly preventing this from gaining traction.
Otherwise, it's Windows Phone OS all over again.
I've been hearing this for over two decades and yet gaming desktops are more mainstream than ever. Certainly the current market is negatively impacting pricing and availability, but those same factors are also pushing their competition just as hard. I'm not even sure cloud streamed games was a profitable idea when prices were sane, but it certainly isn't now.
Smartphones are what's responsible for the death of the internet as a place where knowledgeable, can-do people congregated and built a free utopia together. Once the normie hordes arrived and the marketing / advertising / propaganda / doomscrolling / mindless passive consumption people completely hijacked their brains with things like Instagram, the good old internet promptly died, and was replaced with the VC money-soaked online world we have now.
Gamers don't buy computers anymore (despite their consoles being computers, they are completely uninterested in those capabilities, same as phone users), students don't have a clue about files and directories anymore, and as for programmers... as a lifelong low-level programming enthusiast, I'd better stop there.
The correct answer is to just let your garage door opener be dumb, and buy a RatGDO, which works with any dumb opener including chamberlains: https://ratcloud.llc/
I like to think chamberlain sells their openers at a slightly lower price than they otherwise would, hoping to recover it on subscriptions. So by buying their opener plus a RatGDO, I’m hurting their bottom line just a tiny amount. It makes me feel good at least.
Giving them money in the first place is a horrible idea.
>"Wine experts can't tell" is a myth
It's really not a myth. This has been tested several times (at least once a year) in several different ways. Even the most expert of experts don't really have a clue.
[1]https://www.theguardian.com/lifeandstyle/2013/jun/23/wine-ta...
Except when you look into who counts as a wine expert, it ends up being a bunch of college students and people that took a weekend course.
The task of auditing a business is extremely complicated by virtue of data/information being all over the place. AI had not gotten to a point of where it could outperform humans but in this experiment they find out a group of agents are good enough for this task. I still think there are so many poorly defined tasks that auditors just understand implicitly, eg data discovery; before the audit even starts, that humans will always be needed to initiate and conclude audits just like software projects.
And every year a new human comes along and has to learn all of it again.
Of course ONE AI agent and teaching ONE AI is massivly in an advantage as soon as its as good as the avg person in the job and it blows everyone else out of the water as soon as it gets a tag better.
Or they routinely ask startups to develop proofs of concepts, making it sound it may result in a big contract, and after the startup has proven what they wanted, Amazon just stops talking to the startup and develops their competing product internally.
To me this signifies two things: the low hanging fruit has been plucked and we are now on the slow but steady incremental improvement part of the trajectory and the second is that if you share your corporate secrets with these companies you can count on that being used against you sooner or later and they'll be using tools that you can't buy from them at any price.
[1] https://en.wikipedia.org/wiki/Strategic_lawsuit_against_publ...
Far more effective is for you to listen to different journalists opinions and reporting over time and answer two questions: "Is this person acting with integrity?" and "What are their biases that I need to factor into my analysis". It is much more important to be listening to journalists who have clear ethical standards - even if you disagree with them.
Raw data is only a piece too. Journalism is more than rewording a stolen document. It's following leads, establishing sources, and connecting both public and private data points into a narrative that otherwise wouldn't be discoverable.
I find your gripe attributes a blanket presumption of incompetence and malice on journalists. It reads as condescending.
The reason why this story is distributed across so many publications is to quell concerns of deception by omission or similar. Perhaps all the publications fell for the same trick or are in on the same lie, but this is as close to "compile from source" as you can get with reporting on sensitive, stolen, and secret information.
HN readers may be interested to know a bit about the tech stack. The 15 million or so documents, mostly Outlook mailbox exports, were provided to partners in the consortium via a tool called OpenAleph [1], which is maintained by DARC and based on ElasticSearch and a schema called FollowTheMoney [2], developed specifically for the investigative journalism use case. A lot of the original work on this was done by Friedrich Lindenberg (@pudo), who now runs OpenSanctions [3].
The timing was v. interesting given where we are with LLMs -- it's the first major leak I've worked on where it was possible to leverage local models to do some of the heavy lifting, and this only got more possible as the project went on. The coordinators provided a centrally hosted chatbot with RAG, but I also managed to get some surprisingly useful material using Qwen3.6-35B-A3B on my 24GB M2 MBA via a custom agent harness for Aleph/OpenAleph. [4]
Happy to answer any questions about techy stuff, though I'm not a spokesperson for the project by any means. Seen a few people saying "just release the data!", but I think that slightly misunderstands what this work is. The whole point of forming a consortium of hundreds of journalists and sifting through the dump is to identify the bits that are in the public interest to report on -- especially important in jurisdictions like the UK, where I'm based, which have very strong defamation laws. Think Panama Papers, not Wikileaks.
[1] https://darc.li/openaleph/ [2] https://followthemoney.tech/ [3] https://www.opensanctions.org/ [4] https://github.com/data-desk-eco/sift
> Our reporters will continue to evaluate the material journalistically over the
> coming months. We will not publish the raw data online.
One of my biggest gripes about journalism today is that biased actors will change the narrative to fit their biases via selective omission of facts, thereby putting forth a misleading version. One doesn't have to actually _lie_ to deceive the public; one can just hide extenuating facts and let the public form their own (misguided) opinion.
This is why I feel like the raw data should be released so we can do our own analyses. Why should I blindly trust a random self-selected, possibly anonymous, group of people?
The company eventually collapsed, after revealing that €1.9 billion were missing. The CEO was arrested and the COO is believed to have fled to Russia. The Minister of Finance was asked to testify but nothing came out of that - despite being involved in this and another scandal (Cum-Ex), he nonetheless eventually became Chancellor of Germany.
This is the blueprint for hacking people's first impression.
This would be such a cool feature to have on HN comment threads at the top.
Perhaps I’m one of few, but I typically don’t open the linked article. Instead I come directly to HN comments, read those first, and then decide whether to open the article or not. Would be cool if HN included a brief AI summary at the top.
Forgive me if that’s considered bad etiquette, it’s purely a personal preference of mine, but for what it’s worth I do refrain from making comments myself unless I’ve read the actual article.
It’s tongue in cheek. But shockingly prescient. I am still surprised they didn’t somehow get involved in like FTX’s last—and only last—round.
The linked articles require dedicated time to read, (which is good.) Many people will open this, see that it requires time to read, and move on without understanding the major points of the scandal.
A higher level summary will get the point across much more quickly.
A human is just a person-shaped clump of cells.
I know I'm a bit ridiculous with my over-analysis. I know people say "human" mostly because it's a trope. But it grinds my gears nonetheless.
This thought is likely to be revolting to many people, but most of them are going to compare it to a reality that may not exist: They want infinite human caregivers with infinite patience and infinite money to pay those caregivers, but sadly all of those things are finite.
AI and automation are often second-best solutions to some idealized version of reality, but when the idealized version isn’t attainable pragmatic people start accepting the second-best solution. There are a lot of unpragmatic people who will get very angry at the second-best solution for existing, though.
Compared to the, sadly too common, current hellscape of some underpaid and over worked person who comes by and does the bare minimum (if that) in a rush before hurrying off to their next client.
If only it were that simple to define what a job is and which jobs are “worthy” or not of being reserved to humans.
I’m against dreamy proposals like this one because they miss the forest for the trees. Like previous technologies, AI can do lots of things that humans don’t really care about automating. But deciding which things won’t be subject to AI hinges on personal preferences, class interests, and even market pressures. It’s not something that can be reasoned about on a global scale, and trying to push this agenda might sow even more division among humans than we’d think at first glance.
1. Challenging to get running properly.
2. User experience often leaves much to be desired.
It quickly became clear I was better off spending time and energy on problems directly tied to revenue, not managing operational infrastructure from the ground up.
Most people simply are not so busy that they need someone to manage their calendar. And if you use these things to shop for you, I can virtually guarantee that you will be paying way too much, because vendors are gaming your gaming.
I don't dispute that the frontier LLM companies stole vast amounts of intellectual property ... like many orders of magnitude more IP than people who are spending years in jail for piracy. That's indisputable.
But when you start equating piracy (I'm reading your book without paying for it) to SLAVERY (enforced labor) ... that's just so completely ridiculous that I find it difficult to take anything this guy writes seriously.
I really can’t stand LLM language.
PO came to me and asked if our card would work on their plane. We ship them a prototype and it doesn't work. They ship it back and send along a card that does work in the plane. We test it, and the card is just faster. We get the designer of the data box (works at the prime) on the phone and talk to him. Explain that our card works within timing specs but isn't working in the plane reader. He says something along these lines: "Oh, our reader implements the communication specification just fine, but only for fast cards. When we designed the box we wanted to hit a certain read/write speed, so we went out and bought a bunch and found the fastest cards and wrote our VHDL to talk to those cards." After we collected our jaws from the floor we updated our card to run a little faster and it started working. They had a lot of trouble finding commercial cards that worked.
And another:
One of the hardest parts to find for a PCMCIA card is stamped sheet metal cases. When we made prototypes, we bought case kits where the PCB just dropped in if you followed the PCB outline and used the right 68 pin connector. Around 2020 the companies that made them started to discontinue them, sent out a notice to offer final production batch ordering to do final lifetime buys. I BEGGED a bunch of groups to go order a few thousand, or tens of thousands and just put them in a closet somewhere (they were like $9/each). No one could make the decision in time, they went obsolete. We ended up designing a new machined aluminum case with super tight tolerances and features and it was like 80% of the total $500 cost.
Typical work would look like this:
A group would come to us and say, "We use these cards, but they wear out and we need like 200 more to keep running for 15 more years, we talked to primes about fixing this problem for us but they said the only thing they're willing to do is redesign the entire data handling infrastructure in the plane for $500M." So we'd get some tiny amount of money (like $200k) and redesign the card and make them ~20 prototypes. It'd work great, then they'd spend 3 years working on production contracts to try to get 200 built for $100k. In 3 years, some part would go obsolete, they'd come back and give us $50k to update the design and test, then start the 3 years cycle to buy 200 again. Rinse and repeat.
If they would have just given us like $30k, we already knew which parts were high risk of going obsolete, we could have just bought everything necessary to build 1000 units and dropped them in storage, but federal regulation limited what was possible. SUPER frustrating to have something that solves a problem and watch it fall apart over crappy policy.
It's sad to see this request, knowing that I was part of the team that designed a solution specifically to solve it, knowing it worked and tested it on the platform, but some part probably went obsolete and they never got enough manufactured to solve the problem permanently.
A fun story: Our cards had whatever industrial-spec microSD card we could find (16GB, 32GB, etc.), and we'd internally limit the disk size to whatever the platform needed. Almost every aircraft had some specific number of MB they wanted because all of their CONOPS listed it, and they didn't want to update documentation or software to put a different number there. We had a plane that could only use EXACTLY 128MB, but they were currently using 256MB cards and partitioning them to 128MB. We updated the disk size to present as 128MB so they could skip this step, and they specifically told us to put it back to 256MB because they didn't want to update the commissioning process to remove the partitioning step.
It was rewarding work, and the card we designed was cheap (~$500 each, which is literally nothing for major platforms like this) and 100% US-designed and made (with the exception of ICs manufactured in the usual places). Still, the procedural and cultural practices of these program offices made it a real struggle to get traction and deploy them widely. Government acquisition is just generally designed around funding primes for current development and then paying through the nose for sustainment later. It was hard to design something for widespread sustainment use because by the time you were done and ready for production, a bunch of parts would be obsolete. Then you'd have to rev the design over and over. We were never allowed to stockpile anything, even if it was trivially cheap, to facilitate production later.
This is more of a survey to identify _potential_ suppliers of NEW CF/PC memory cards. It's not a request for bids, nor asking civilians to root around their junk drawers for cards to send off to the military.
Reading between the lines, they are likely evaluating multiple options for modernizing whatever F-22 system this is, and need to find out if going this route is even feasible.
Obviously you're entitled with this opinion, but others are entitled with their opinions too. For example, they might want to:
> talk about pricing, profitability, and capability
"I don't want any of this" doesn't really add much to the discussion. I can copy & paste "I don't want any of this" to any HN submission, be it a Nintendo switch emulator, the 1001th javascript framework or yet another Lean-verified math proof of an open math problem.
It says this is available to "Eligible Google AI Ultra subscribers" but it looks like my workspace account isn't allowed one of those, so I'm looking with my non workspace gmail account and that says Ultra gets Genie 3. I don't know if that's because I'm in the UK (a previous Genie version was US only) or if its because they haven't updated their marketing spiel and I have no way to know if that would make me an "eligible" subscriber or not.
DeepMind do such awesome work and its a shame that there's this layer of bigcorp frustration for users to get at it.
Being able to spin 360° degrees and not have details around you change...