Sounds like “according John McAfee the threat of malware has been largely solved”.
Edit: it's way worse than that: the actual figures says that Mythos “only” falls for prompt injection 2.6% of the times. Maybe for an ML engineer used to work with unreliable tools that sounds impressive, but for security purpose having a system that fails every 40 attempts is outright catastrophic. Imagine if your OS vendor only patched known security holes in a way that still let an attacker go through every 40 attempts…
And we're just talking about known kind of prompt injection that are tracked by benchmarks, not any kind of 0-day vulnerability found by clever attackers.
"largely solved" as in they the models they trained don't fall for prompt injections as often but not "largely solved" as in the underlying issue is solved at all.
And each day a model remains unchanged, attackers get to test and experiment with how to make it tick. And every attack that works remains on the internet, forever, just waiting to be ingested as context.
"We'll just stop training and we'll be profitable!"
Nah, there have been no improvements in the fundamental issue that makes LLMs vulnerable to prompt injection - data and command intermingling. They may be better at detecting injections today, but that’s an arms race. Specifically, it’s an arms race where as soon as the pace of LLM development slows down the attackers will have a huge advantage.
Incorrect. If every sentence you hear is indistinguishable from your inner monologue and you can’t tell the difference between your uncle saying something and thinking it yourself, seek help.
As someone who just got out of a meeting demonstrating how Copilot running Luna can be breadcrumbed by a single line of text in innocuous package into downloading and installing malware I think Boris Cherny might be wrong.
Boris is talking about Anthropic models only of course. In the chart in his tweet Luna has like a 43% chance of being compromised with sufficient attempts. But even the Anthropic models have as high as a 5% chance, which is still far too much IMO.
That’s fair and to their credit (and somewhat the user’s detriment) I think the user focused guardrails they’ve imposed have unintentionally helped to protect against the kind of attacks I have success with.
But you’re right…5 percent is better than what it was but at scale that’s still quite a bit.
I'm a bit worried on the UX on mobile, so I did some testing to get at least the basic working! Video also worked better than expected, I love how the video background in the Orbital deck turned out! Thanks for the feedback!
Have you considered moving to open source / Chinese models?
If gemini-2.5-flash-lite is good enough for your application, you will find even lower cost options with better performance outside of the Google ecosystem.
I can't, my main task for it was translation, and no other models in the world can keep up with the multilang support like gemini models, because, no other company in the world has the multilang dataset like google and there might never be one, because google has been collected the internet data for it search engine since the 90s already. And so much has been lost in the history of internet and I doubt there any other company in the world that can go back in time to get those data back.
I worked in a startup whose value proposition was largely derived from using graph data and graph databases under the hood. The main benefit is, as the repo even states "Fast data processing: Links traversal is processed with O(1) complexity."
So, technically, you can do deep traversals quicker. A few notes:
1. Few use cases truly need low-latency deep traversal on realtime data (>5 hops). There are some well known ones like fraud detection in payment processing and, possibly, social media recommendation engines. But I am not even sure how latest social media engines work, and whether they still rely on graph DBs.
2. However, in practice the advantage is often marginal. With modern analytical databases, or even an optimized PostgreSQL (ltree, materialized views, pgrouting, pg_duckdb, etc), you often get more than good enough performance. In addition those traditional SQL DBs scale with hardware more easily than graph databases. So, you can always use the lever: "Throw more hardware at it."
3. Even Graph DBs don't get good traversal performance under all conditions without hand tuning. For example, there is the "super node" issue, a node with an abnormally high number of connections (edges). And if you still need hand-tuning, you might as well choose something more versatile.
4. The ecosystem of a PostgreSQL and other popular DBs is just unbeaten. With graph DBs, you often prematurely put yourself in a corner that you don't want to be in.
Hence, my recommendation. Unless you are really sure that a graph DB is the right fit for your use case, start with something else, and go the graph db way when you have established a true need.
It is unfortunate for the employees but it is necessary, and hopefully also a wake up call for other German companies and the government.
It’s time for leaner structures, smaller teams, faster decisions, and shorter development cycles. Volkswagen, like no other company, symbolizes the lethargy of the German economy and business environment.
Hopefully, the seven years with barely any economic growth will end soon.
None of that will address the fact that China is intentionally engaging in mercantilism and keeping the value of the yuan low to maintain export volumes. cutting 100k jobs will shave 1k off the production cost of a car, china has like a 6k advantage.
> Volkswagen, like no other company, symbolizes the lethargy of the German economy and business environment
A car development cycle is/was seven years in Germany. And that was totally fine. When the "Golf 7" comes up, the "Golf 8" doesn't come out only 2 years after.
What happened is the EU (and Germany in particular) left only two development cycles to the biggest industry in Germany (and hence the biggest industry in Europe) to entirely change.
For the EU said: "In 14 years you're forbidden to sell a single brand new ICE car in the EU" (of course they recently changed that tune, but too little, too late, the damage is done).
But that's not all: not happy to drown the german car industry, they decided that the corpse wasn't drowning quickly enough and Germany decided to shut down all their nuclear reactors (I think they recently reversed their stance on that too) and become depending on Russia as a major energy source (which the US warned Europe / Germany to not do).
Of course the shit hit the fan: Russia attacked Ukraine and everything turned to shit.
And now it's estimated German car makers pay their energy price 7x more than was chinese car makers pay their energy.
So, no, it's not all "the lethargy of German economy and business". Germany was (and to some extent still is) producing incredibly fine driving machines: not just VW but Audi, Porsche, BMW and Mercedes.
And those cars did what was asked of cars: drive very finely.
It's eurocrats/bureaucrats and german politicians who killed the german car industry.
P.S: I've got nothing against EVs... But don't be an ideological idiot about it: don't kill your main industry while handing the keys of the kingdown to China. The shift was way too quick. The EU should have been way more cautious and shouldn't have stupidly closed so many nuclear plants (to moreover start burning more coal again).
> What happened is the EU (and Germany in particular) left only two development cycles to the biggest industry in Germany (and hence the biggest industry in Europe) to entirely change.
Thats of course complete BS, the change to EVs (and more economical ICEs and Hybrids in between) could be seen decades away! And VW knew it, they just chose to ignore it and continue down the easy (lazy) path.
There is nobody to blame other than VW management, the Piechs and Porsches in particular.
EU did the Green Deal, though, which is way worse than shutting down a few nuclear reactors that after all didn't produce that much energy anymore and were anyways planned to be shut down.
We pollute right outside our block, give credits to companies claiming to be "green" so that we can sell hardware here that is certified, so that we can go to sleep a bit happier as human beings, because we believe we're doing the right thing.
In all this, they make tons of money, make our economy worse than it ever was with no end game.
You drive your Audi, BMW, Porsche. Mercedes. I'm gonna drive a Toyota or a Honda since I don't like costly repairs, subscription chair heaters or getting screwed over.
You are in for a surprise if you think latest generation Toyotas or Hondas will be less costly to repair. Just look at the recent debacle with the V35A-FTS.
For Europe to meet its climate goals that was necessary. And BYD seems to have no issues shifting. Why should VW?
Though I do see another option emerging: people just not buying cars at all. More and more friends are relying on public transit alone. That's an even better option for the environment and climate of course.
For the economy it's not so good but it'll have to cope. If people don't buy cars it means they'll have lots more money to spend on other things anyway.
Both are part of VW Group and have seen marketshare fall as well due to tariffs as well as China cutting down on foreign luxury brands in favor of domestic ones.
Barely 5 years ago a black Audi A8 was the go-to VIP car in China.
> BMW and Mercedes (Daimler Grouo)
They will survive, but they also made sure to diversify beyond China and the US, which VW never did which is why they are so screwed.
> I've got nothing against EVs... But don't be an ideological idiot about it: don't kill your main industry while handing the keys of the kingdown to China. The shift was way too quick...
Also
1. Don't have half of your production depend on the mercy of two SOEs (FAW and SAIC) who are trying to build indigenous competitors.
2. Diversify markets - America and China don't want to buy "Made in Europe" cars anymore. Mercedes and BMW made sure they were status symbols in multiple markets the same way VW's Audi and Porsche never could.
3. Don't let other countries have veto power within your largest conglomerates - Qatar has veto power and 17% of all voting shares in Volkswagen Group
Why does this obvious falsehood get repeated so much on hn Everyone can look up what the CDU did, it's not like they did it in secret.
It's bad enough that nuclear proponents ignore the massive cost of nuclear compared to renewables once you factor in building costs, insurance and storage imho. But then to pretend like the conservatives didn't have a choice but to bungle the nuclear exit really is too much.
The CDU never listened to the greens when it came to not killing the solar industry and serving it to China on a golden platter. Or not killing our rail infrastructure by continually delaying maintenance so that we would have to do much more costly repairs later.
But in this instance, they do, and do it in a way that is a gift to the energy companies, and instead of noticing the obvious corruption at play, people still blame the greens.
Sure, SPD didn't have anything to do with it. Schröder and the Greens didn't have anything to do with it at all.
The Green party itself was literally born out of a protest against nuclear reactors in the 70's (Whyl).
CDU tried to slow down the phase out, but when Fukishima happened Merkel thought about staying in power rather than doing the right thing (= not shutting down the power plants).
I don't care about CDU, which as usual never misses a chance to disappoint, but not saying that the phase out was planned and executed by the Greens (and obviously SPD) is just factually wrong, because they initiated it all and implemented it. It wasn't just a draft, "let's see if it gets amended, etc..".
EDIT: I was talking just purely about energy and such reforms. The last 30 years have been ... to say the least... "weird" and short term.
Disabling a technology is a huge mistake, no matter what. No country did it the way it was done in Germany.
Yeah it's crazy, in every post about Germany inevitably the nuclear phase out will get brought up and the comments/discussions always seem like a shit show of misinformation to me. And it's always how nuclear is being phased out in favor of coal and gas and that's why energy is expensive, industry is dying, and the Greens and or EU are to blame for all of it. It's tiring, really, and I have the impression that these posters are not acting in good faith, but posting to reproduce a narrative.
As we are supposed to strongman other commenters arguments here, I'd say they are not posting to produce a narrative, but rather that they have allowed themselves to believe that they are the "obviously right", and everyone who believes otherwise is driven to do so by ideology instead of logic.
By painting the opposition to nuclear energy as emotionally driven, they can assure themselves that their position is rational, and this allows them to subconsciously ignore the evidence against it.
your misrepresentation of the timeline makes your comment meaningless. Germany has been winding down nuclear reactors for 20 years.
the only actual story here is that the market trend (both tesla in the US and all chinese EVs) pointed towards EVs being the future and German/Japanese brands have not been willing to give up their cash cows to EVs. The brand value will remain (e.g. germany MIGHT be ok), but the value and reliability differentiators will disappear.
> Of course the shit hit the fan: Russia attacked Ukraine and everything turned to shit.
When Russia attacked Ukraine in 2014 absolutely nothing happened. Covid showed that Europe is eating boogers with their pants down, so both Trump and Putin decided to go on the rampage.
> A small task that cost 121,000 tokens done directly cost 513,000 tokens when fanned out to two subagents, because every subagent has its own bootstrap cost, and the parent then consumes its transcript.
Is that true? My understanding is that the subagent only returns the result of the request without the main agent consuming its entire transcript.
Years ago I worked on an event system that correlated about 500k daily events to about 40 actionable daily events.
A big part of it used prolog to map artifacts to application to business and technical accountable individuals. So if a down storage device offlined a database and broke an app, the business user and storage guy would be called or paged.
My team does this with Splunk today. For probably 50x the compute and 10x the cost.
I see. Cool. So did you generate the inference / reasoning rules to map those 500k events to 40 by hand or are there solutions to automatically discover those rules based on past incidences?
We wired up the network monitoring systems which built out the hierarchy of network gear, then used a fairly lightweight filter/rules engine to dedupe and normalize events. For example a Cisco 6500 switch might throw 100 events when an interfere dropped. We could roll 90% of them with the filter. Another device would send a junk “interface down” alert periodically… except an attribute would say “is_down=false” lol
So we pulled in our business-artifact mapping system (this would be ServiceNow today), the on-call rosters, the network topology, runbook/kb, and some other goodies and grabbed the attention of the right person at the right time with specific guidance about what to do.
Basically if a switch, server or critical app failed, we immediately knew what system was impacted, the scope of the impact, who to inform and who to call to resolve. Eventually we expanded it to batch non-critical failures and schedule repairs during outage windows and identify specific dev teams for components of larger apps.
I left after that. It was a fun project and a big break for me, all because I was the only person who had heard of prolog in a happy hour conversation!
> are there solutions to automatically discover those rules based on past incidences?
While it's not exactly what he did, the answer to this question is yes. That's called inductive logic programming. Essentially you provide background knowledge, positive examples and negative examples, and it spits out a logic program. It's where the frontier of symbolic AI has sat for a hot second now.
Thanks! There is something magical about those symbolic approaches. But in practice, they seem to often fall short of expectations. Wonder whether there will be another "symbolic AI summer" one day.
"Halle is in the state of Saxony-Anhalt, which has an election in September. The far-right anti-immigrant Alternative für Deutschland (AfD), which is leading the polls with about 42%, has seized on the row."
[1]: https://x.com/bcherny/status/2086520950259118464