I'm good with DeepSeek v4.1 set to high. It is a relentlessly "hardworking" dirt cheap model.
Told it to convert a products page (that had two different fonts based on language) from two columns layout to 5 columns on desktop and 2 columns on mobile ensuring typography is readable.
My man went into spawning sub agent which failed to drive chrome so it wrote its own chrome driver protocol server in Typescript then generated a prototype website then downloaded the images and rendered each variation in a directory taking 100+ screenshots analyzing the typography depth and then delivering detailed report and then writing the whole thing with new page layout testing it again with several dozen screenshots using its driver and then saying all good and all really was good and whole thing took 25 minutes or so (including double visual validation) because it generates token at an incredible speed.
Total cost of the above? $0.07 cents.
PS: It generates token at such a blazing fast speed that you can't recognize the words as they are being added and can't read it without scrolling and pausing even if you're Jimmy Carter.
"Ah, yes. This is a classic dog-breed-to-appliance-failure mapping problem."
ChatGPT 6 Pro answered it without issue.
Isn't this basically the model admitting it was trained on this? Otherwise why would it think a pelican svg is a usual request?
All four levels have a correctly shaped bicycle frame. The differences between the pelicans aren't huge, but the xhigh one has a better beak.
I haven't managed to get one for level "max" yet, it hit the limit of 128,000 cap for output tokens while it was still reasoning about the question!
Max started its thinking trace like this:
> This is a classic test request, so I want to plan out a well-composed pelican with its distinctive beak and pouch riding a bicycle with proper wheels, frame, and pedals, set against a simple sky and ground backdrop.
So that failed attempt on max cost me $2.56.
I ran this using my llm-anthropic plugin:
uv tool install llm
llm install llm-anthropic --upgrade
llm keys set anthropic
# paste key here
llm -m claude-opus-5.5 -o thinking_effort low "Generate an SVG of a pelican riding a bicycle"
# Then to save the markdown logs
llm logs -cu > logs-with-usage.mdI hope Opus 5.5 is better, if for no other reason than all the Claude slop I have to read will be at least more tolerable.
One funny side effect of all of this: realizing that coworkers that use AI for almost all the text they generate at work have their writing style change every time a new model ships.
The user is right. The outage is a real concern, and the issue is worse than we realized. Requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5 encountered elevated error rates. Worth stating plainly: these are not just models — they are load bearing rungs on the software development tooling ladder, and a blocker on this level makes the outage really bite.
One decision that is yours to make, not mine: should an email be drafted to Anthropic support? This issue has teeth, and a canonical handoff can land us where the main gate is no longer breaking silently.
I think this is what I'm most interested in. I mostly moved to Astra because I just can't work all day with the Claude Opus 5/Fable writing style. I don't think Astra is a better model, but it's the first OpenAI one that seemed good enough to me. Definitely keen to try Opus 5.5 and see if this claim is real.
I'm on the fence about calling out AI-isms but I think it's definitely worthwhile to call out ones that actually don't make sense.
Edit: In response to the initial replies. To me it clearly means "releasing frontier models at any pace less than as fast as possible". It implies relative restraint compared to the previous state and without stating the degree of restraint.
Nice. I was starting to think that Haiku got abandoned.
Interesting how the very first line is used to remind the reader of their call to pace the frontier just last week, and everything else after that line is to demonstrate with very specific numbers how they absolutely are not pacing.
I could see them accomplishing it and seeing gains like this in roughly the correct timeframe, and when I heard about that development I assumed the frontiers would probably jump on it.
How it works: https://miraflow.ai/blog/deepseek-v4-1-flash-causal-encoder-...
Ah, they're spreading their limits to all their models it seems. Definitely not a good thing long term in my opinion.
Prices per 1M tokens Claude Opus 5.5 Claude Opus 5
Cache reads $0.20 $0.50
Input tokens $4 $5
Output tokens $20 $25
Cache writes $5 $6.25
If you are forced to reduce price despite raising capabilities, that certainly tells something about the market, and potentially about Anthropic future profitability too, since this model is their biggest topline contributor
God I hope so
> Input and output tokens are $4 and $20 per million, 20% less than Opus 5. Cache reads (which make up the majority of agentic and coding work costs) are $0.20 per million tokens, 60% less than Opus 5. Opus 5.5 also generates output more than 30% faster than Opus 5.
Better than Fable, cheaper than even the last Opus. I use Opus as my main driver so this is very exciting!
Sounds like they noticed the complaints. I'm curious to see what LLM-isms this one may have.
I work in secure systems and it’s shocking how many people believe this - the incentives from management are all about it too.
They should try something like "100 agents at noon on Sep 23 do the chicken dance for 30 min in the middle of the street in DC, then we'll consider not releasing the info and not sell it to the Chinese".
Of course there is. For example, SeL4’s security and reliability proofs still hold in the world of LLMs. The problem is that most software isn’t written on that firm foundation. Instead, most software is made by people with the philosophy of “if it looks like it works, ship it”. You don’t get secure software by working like that, because security vulnerabilities aren’t visible.
We - humans - know how to write secure software. Just like we know how to make safe aeroplanes. The problem isn’t that we lack the capability to make secure computers. The problem is we don’t have a culture of security. Secure software is - somehow - niche. And as such, it’s much more expensive. And nobody wants to pay.
That is exactly the canon.
For example, do not hook your goddamn water or traffic or electricity infrastructure up to the goddamn Internet, and then, do fire the guy who suggested it.
The correct analogy for computer security is not locks and keys and doors and gates. It is a house in a floodplain. Your house will not survive the flood of it hits you. Do not store anything critical or irreplaceable in that house.
In the pilot, the Galactica was literally being turned into a museum, and that's why they lived.
China hacked 22.1 million records of US government employees:
https://en.wikipedia.org/wiki/2015_Office_of_Personnel_Manag...
Perhaps firing expertise and hiring incompetents wasn't a good idea.
Not a skin-scientist, but this seems misleading: Sunburn is not caused by some bad DNA floating around, nor damaged cells making making mistakes or pumping out wrong proteins to inflame their neighbors.
Instead, your cells have techniques to measure a variety of risk factors and signals (which does include DNA-molecule weirdness [0]) and when things get too risky the well-behaved ones will self-destruct, to avoid the chance of turning into a cancer.
So the sunburn you see and feel is instead due to cells signalling "something damaged me and I can't be trusted, I'm shutting down, recycle me" to your immune-system, plus the attendant bustle of hazmat-cleanup and increased security-sweeps for cancer-leftovers and other potential causes of the damage.
We wrote a research paper on the general principle a few years ago: https://conferences.sigcomm.org/hotnets/2022/papers/hotnets2...
According to Obscura's legal page, it is a New York-based company [0]. Under US law, a secretive court order could compel a US company to update software or implement targeted logging on a specific user without notifying that user.
The only scenario where Obscura would be useful is if Mullvad were compromised. Why would I trust a New York company to shield me from a more reputable Swedish company?
[0]: "(2) your written notification must be mailed to 169 Madison Ave.; Ste. 11185 PMB 63183; New York, NY 10016..." https://obscura.com/legal/
Many (if not all) of the benefits on the landing page are available in Mullvad too, which is a more mature and reputable product, has all clients fully open-source, and powers the exit servers for Obscura.
Why should I choose this over Mullvad?
I obviously am I biased, and think this is very very cool. There are many intangibles here beyond the budget numbers; a bit of history, a bit of whimsy, a bit of color to what is often utilitarian (public transit). It certainly attracts tourists to the city (I was pretty surprised at the number of out of towners at Muni heritage weekend).
But I do think it's important, not just for tourism but also as a legitimate connector along the Embarcadero corridor that otherwise has no public transit along it. Agree that a breakout of the cost of maintaining these things would be appreciated, not sure if that exists in clear terms, but I don't think it's massive. For example Car 162 was damaged in a collision in 2014, and took 12 years to fix and return to service. For many gripmen / Muni restorationists working for the city, it's a side project.
And while I'm here, I'd be remiss if I didn't post an article about Maurice Klebolt, one of the biggest advocates for vintage transit in San Francisco. It describes how he acquired car 106 from the Soviet Union [2]
[1]: https://www.streetcar.org/streetcars/162-162-muni-1950s-live...
I installed ChatGPT on her laptop and allowed it to run a certain subset of shell commands and it finally worked.
If im gonna have to muck around in the terminal to do the most basic of tasks, I may as well use Linux so I actually have a semblance of ownership on my computer.
I’ve been dabbling in Linux since the early 2000s with Yellow Dog, SUSE, and Ubuntu, tried making it my daily driver in the early-mid 2010s, and for the past 5 or so years have been using it in a secondary/tertiary role.
Recently I’ve been trying to move my gaming box from Windows to Fedora (which I’ve found mostly makes the right tradeoffs to suit my preferences). I’ve gotten things to mostly work, including games that have been blockers in the past, but just yesterday my ethernet connection listing disappeared and broke network connectivity. No big deal I thought, just go into settings and add it back, but as it turns out the KDE settings UI for NetworkManager is horrifically flaky and most critically, its add interface wizard doesn’t set correct settings for a successful connection. No, you have to drop down and use nmcli afterwards to fix things because the UI also doesn’t surface the right checkbox. I wouldn’t have even figured that out if it hadn’t been for a chatbot’s assistance.
This is the exact sort of time-burning nonsense that isn’t a problem on either of the commercial platforms that I tend to encounter under Linux. Is this single instance going to push me to reboot that box back into Windows? No, but if issues like that keep cropping up eventually one will become the straw that breaks the camel’s back. Desire to avoid continued frustration will win over the logical portion of most peoples’ brains.
That's defined as "the device has become a brick, cannot be recovered". It was mostly popularized by broken firmware updates which effectively destroyed the device, as it could no longer be salvaged.
You calling a borked package install in userspace almost bricked is telling about your technical capability.
And I agree that people that are basically technologically illiterate should not use Linux.
They could if someone administers it for them, but if nobody can help them - it's just not the right choice.
Compared to the last XPS I had, it's substantially better for my use case, since I'll generally be remoting into systems anyway.
It's placed right next to a similar report for the user to see which apps generally connect to the internet and to which domains.
Basically, the main point of the article, "I discovered that macOS 15.3 had enabled a feature that was phoning home every 15 minutes with personal data" is not true.
Some people understand the concept of trade-offs but don’t understand how other people don’t want the same trade-offs as they do
If you're not willing to stop using harmful software that's your problem.
And let's not forget that cheating in online games is only an issue because companies have co-opted what was a social activity between people you know (or at least can get to know) into a more profitable asocial e-peen measuring contest between randoms.
If you look at MacOS with a critical eye, it's nearly impossible to say it has good UX. It's outdated and everything requires more clicks. Finder stuck 20 years ago. Hard to find settings. Poor screen real estate. Hard to read text with transparent everything.
Let's kill this myth that "Macs are more intuitive to use".
Yes, there is a whole moral discussion to be had, but as much as you win points for doing the better thing, the better things don't really pay bills.
The good news though is that we definitely won the server war. We can spin up VMs and install Linux without licences. We have access to software that works well, security updates, frequent releases, a massive ecosystem of free software, no advertising, very little to no licensing for software, and it works incredibly well. So much so, that we can't even imagine a world in which servers require licence keys or subscriptions or a marketplace account.
Imagine now, if laptops & workstations were as solved as servers are.
That's why we aren't there yet, because they're not.
But don't fool yourself. You're subverted to Apple's rules the same as Microsoft users are to Windows.
I prefer to fight a little bit on an update on a unix system on something that is not working 100% to give up choices. When you use Windows, MacOS, Android you do not have choices. Choices are dictated from above, like it or not.
As for me every time I have to interact with something that is not BSD or Linux I suffer miserably.
Build quality (aluminum), charging (magsafe), screen resolution, battery life, noise (or lack thereof), trackpad…
It’s like other vendors aren’t even trying.
Also the OS mostly “just works”, just last week my Linux laptop disabled NVIDIA GPU (and almost bricked itself?? Not sure, had to fix apt) during automated updates
I feel this pattern is really abusive. Like I am offering to slap you in the face, giving you two answer options: Yes and "Maybe later/no, thank you". As though you have any obligation to be forced to answer this question again at some late time, or to be very polite while you refuse to be slapped in the face. The correct answer is obviously "No, fuck off".
It's always "maybe later" and never "no seriously, I don't want this, go away"
I know there's been discussion about whether pelicanmaxxing is happening, but this is at least evidence that Claude was explicitly exposed to this problem.
I've failed twice to get "Generate an SVG of a pelican riding a bicycle" to work with max, because in both cases it ran out of the 128,000 token budget while it was still reasoning about the problem.
I'm suspicious that "max" may be virtually useless if it's that easy to have it overthink to the point that it doesn't get to a response.
Transcript for one attempt here - expand the "Reasoning trace" bit to see it: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Edit: https://artificialanalysis.ai/models/claude-opus-5-5?models=...
> Paul Ryan 9 years ago
> Note that locate_template() does not prevent directory traversal attacks, so if you’re passing a user-provided template name to the function, be sure to verify that it’s from one of the three appropriate locations (active theme directory, parent theme directory, or /wp-includes/theme-compat/ directory).
https://developer.wordpress.org/reference/functions/locate_t...
I guess it depends on what you're used to? The last time I tried using GNOME, I found it very unintuitive. I ran into stuff like:
- Why don't windows have maximize and minimize buttons?
- Why do I have to go to a separate screen to un-minimize a window?
- Why does launching software require going through the GNOME equivalent of the Windows 8 start screen?
- Why can't I make the trackpad scroll speed slower?
- Why don't I have any system tray icons?
- Why do the extensions I installed to work around some of the above problems (some of which have millions of downloads yet haven't been added as settings) break after every GNOME update?
But the Apple with ads is not the Apple that had some taste and discernment in the past. For a long time I’ve visited the App Store’s app update page directly (tap and hold on App Store icon to see the context menu option). Anytime I inadvertently go to the App Store home page or the few times I search, it’s an ad filled disaster!
From this article
> repeatedly attempting to prod customers towards even more of the company’s products might seem cheap, even distasteful.
From a recent post by John Gruber:
> Steve Jobs in 2011: 'We Build Products That We Want for Ourselves, Too, and We Just Don't Want Ads' [1]
Looks like Tim Cook, John Ternus and Eddy Cue really enjoy being swamped with ads in their products. Will there soon be a time when Apple executives start carrying some other brand’s devices with them to avoid having a rotten experience?
[1]: We don’t want ads https://daringfireball.net/linked/2026/07/28/jobs-we-dont-wa...
Recently I am starting to question some of the design decisions. Most native apps are bloated with functionalities, most of the time useless. Ads pop up everywhere, even before persistence; iCloud and backups are clearly designed to make you pay, the free space is useless.
I am not that sure my next phone will be an iPhone.
I have tried a recent Mac, and the system isn't intuitive at all. The first time I used GNOME, it was a breeze. Not the same with the MacBook system.
But it looks like the rdrand16 instruction can produce zeros just fine, it just sets CF=0 erroneously (indicating an error and that the user program should retry).
So keep that in mind when you try to reproduce it too and use some abstraction that could implement retries internally.
return 4 # Determined by fair dice roll.Do we now learn that they fixed "always generate all 1s" with "never generate all 0s"??
EDIT: I've been unable to reproduce the problem on my CPU, FWIW. It's a Ryzen 5 3600.
EDIT2: OK, update, I can reproduce it with rdrand16, rdrand32 is fine but rdrand16 can never generate all 0s. So my CPU does have this problem!
You definitely would expect a roughly equal number of 0s as any other of those numbers since it's uniformly distributed. And definitely not 0
(I see a RomCom script where the chatbot decides to get two people together, and acts as Cyrano de Bergerac for the handsome-but-lunkheaded farmboy, and bestie/confidant for the girl-about-town).
The difference is in consistency in enforcement and general background crime rate.
If you have a nearly guaranteed chance of getting caught doing something, the punishment almost doesn't matter. If every single time you went over the speed limit you got a $500 fine 100% guaranteed, there would be effectively zero speeders ever. If you have a 1 in a million chance of getting caught but the punishment is 10 years in prison - there will still be plenty of speeders all over the place.
Humans do not do well with estimating risk on low probability events. We are all just big children in the end. Consistency in consequences are what actually matters, not overly harsh punishment.
First hand experience here - worked with a non-profit that has a tiny solar and battery installation for a light display, and we have had constant theft and vandalism problems with it. It's not even easy to get to - those meth-heads had to work to steal the lead acid batteries and copper wire...
No because now you're using land that could be used for other things, which is exactly what this avoids. This takes existing land that loses water and costs money, and makes it lose less water and make money.
Nothing short of a face-to-face security response will deter these criminals, so there it is a challenging cost-benefit balance. At least panels in parking lots and other developed spaces have the benefit of witnesses. I'm sure a 20 foot climb up a steel beam over running water is a decent deterrent as well.
Put the solar panels in a field: The solar array uses less copper. The shade supports don't have to hold up solar panels: Shade supports cost less.
The best reasoning they give is that California has insane permitting requirements, and it takes 1/6 the time to build on developed land compared to undeveloped land.
Recently it had TOTP 'MFA' added to comply with a Secretary of State mandate.
Anyone who uses the software can just open the database files directly. They're just DBF files in a shared folder on a file server. All the users have to have read/write access to the files or the application won't work.
I hang my head.
My recommendation is to get rid of the DBF/DBC files and move to a SQL DB of some flavor ASAP. If you have the source code, use ODBC or OLE DB to point to a server.
Source: filed that bug over 20 years ago when I worked on the Fox team. No, it wasn’t going to get fixed without rewriting large parts of how the DB engine worked.
However you sourced your target is irrelevant to the activity at hand. When I first heard of this news I knew immediately someone skipped target verification or that it was simply no longer a policy of the DOD.
There needs to be some protection for first responders and others helping the injured.
> The Pentagon chief last year slashed offices that didn’t contribute to his goal of “lethality,” including the group that assists in limiting risk to civilians, known as the Civilian Protection Center of Excellence. Around 200 employees who worked on the issue, including at that office, have been reduced by about 90 percent, according to two current and former officials and a person familiar with the effort. The team that handles civilian casualties at Central Command, which oversees the Middle East, has dropped from 10 to one.
Negligent homicide in pursuit of their DEI goals, basically.
This is what happens when people who don't understand AI and its limitations embrace it wholeheartedly, thinking it's going to behave like The ultimate analyst. Except it's not. Its got blind spots in more places than anyone knows. And not to mention the system itself has multiple areas of failure.
I feel like we're optimizing the wrong metric here.
Reading the details, "AI" doesn't really seem like the culprit -it's a scapegoat.
The intelligence that it was no longer a military target never entered the target database, the team that was responsible for vetting the target list was gutted, and said team was never even consulted.
The White House wanted 1000 targets and pulled from their database without any due diligence. Whether it was an AI call or an SQL query - this was from pure human maliciousness and incompetence.
I cannot believe this has come to pass.
Also, moat discussion is the lowest form of discussion. I don’t care if jev has a moat. Did it get the interface right? What other past ideas have we overlooked that if given some love, could kick the door down like jev did?
Really silly stuff.. people wanting to talk about moats when there’s no castle. Moat talk merely projects the illusion of being engaged but, much more often than not, it’s hollow engagement.
- For general chat and web search, occasional image editing, small coding work, document review etc. ChatGPT Plus is basically limitless and “just works” since 5.6. I’ve yet to give it some task it cannot do.
- When given sensible instructions, it hardly annoys with weird phrasing, glazing, or annoying constructs.
- The apps are very good (ignoring the initially terrible Codex app)
It’s easily my best spent $23 a month.
Cached Read: ~6,500M
Input: ~150M
Output: ~20M
Approx $40 worth of usage across DeepSeek V4 Flash + MuseSpark Contributor 1.3. And a bit of both the GLM models. This is covered in a $10 subscription.
If I were to use Luna's API pricing:
$0.02 x 6,500 = $130
$0.20 x 150 = $30
$1.20 x 20 = $24
So $184. And this is assuming smaller coding sessions (<272K) beyond which Luna pricing doubles.
--
Cost wise, these models are nice for small stuff. Translations etc. Any model that does not provide multiple Mtoks of cached reads per cent is not very useful to me for coding workflows.
They're both pretty horrible, but I find it difficult to find arguments for why Anthropic is worse than OpenAI, other than their doomtrolling. Which, in the grand scheme of things, doesn't even register.
Edit: forgot about the SpaceX thing.
I dont know how they make money here
Snark aside, Luna 5.6 was (is) an incredible game-changer.
This hasn't been the case since around July. If you measure usage in raw api costs, Anthropic is actually giving more on $200 than OpenAI now. This includes resets. Usage allocation difference would be humiliating for codex subs were it not for resets. But fixing usage limits with resets is ugly, and they're not good for your mental well-being.
> Context window in the harness
Codex now allows 1M for subs with config params. But generally speaking, you shouldn't really be using 1M context. If you accidentally send a request with say, ~700k context already accumulated in a session which is outside cache TTL, you're paying full cost of these 700k tokens.
> I've subscription hopped a bunch
OpenAI actually has a new strategy to prevent subscription hopping after their 2-3 month-long marketing push to get claude-folks to switch over:
you can't buy a $200 sub anymore. So if you cancel, you won't be able to get back in. Hostage situation, essentially.
EDIT: re: usage limits, oh-my-pi maintainer has been tracking this - https://nitter.xitter.cc/_can1357/status/2090075496948060372
Here's GPT-6 Luna pelicans: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
And GPT-6 Sol: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Scroll to the bottom for the GPT-6 Sol max one: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
For comparison, here are the pelicans I got for GPT-6 Astra: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... - I still like the Astra Max one best.
Here's a comparison grid showing all of the GPT-6 and GPT-5.6 pelicans at all effort levels: https://static.simonwillison.net/static/2026/gpt-6-and-5.6.h...
The grid is actually really interesting, because it shows that the 5.6 family default to brighter colors than the 6 family.
I actually preferred 5.6-Terra not because it is technically superior (it isn't) but because it had better instincts to NOT do this stuff.
PS - Speaking of better instincts, have they closed the UI-design gap at all? I keep a Claude subscription just because /design produces significantly higher quality UI design/UI feedback/UI refinement than anything I've seen from OpenAI.
1/ Usage limits: downstream of input/output cost, but resets and obscure windows and odd 20x plan / 5x plan != 4x usage math throw a wrench into it. Winner right now is Codex by a mile, especially when you factor in the fact that ChatGPT usage (even 6 Astra Pro) is essentially unmetered on the 20x plan. Always a bummer when asking if I should see a doctor about a rash means I can't code as much. It's also is a godsend if you use an MCP like oracle to automate the process of calling the Pro model on particularly tough problems, giving better planning results or deeper code analysis without burning usage.
2/ Context window in the harness. Claude Code wins on this. There used to be a toml file workaround for Codex to extend the GPT context window to 1m, but this stopped working on the plans and only on per-token billing (ETA: noname120 pointed out this is no longer the case and it can be enabled again [1]). 252k is just not enough. Codex's compaction is very good, fwiw, but it happens so frequently that even a model as powerful as Astra sometimes loses the plot on long-running tasks.
3/ Ability to use the plan outside of the official harness. Codex wins. Anthropic does shit like bills requests as extra usage if it sees a hermes.md in a commit.
I've subscription hopped a bunch, and at times I've had both, but I keep coming back to Codex because it wins on 2/3.
ETA: apparently I haven't been Keeping Up With the Altmans and new 20x signups have been disabled for a few weeks. I am grandfathered in, which makes the comparison above pretty much moot.
Goodbye Hacker News--it has been a good run.
Prices per 1M tokens Claude Opus 5.5 Claude Opus 5
Cache reads $0.20 $0.50
Input tokens $4 $5
Output tokens $20 $25
Cache writes $5 $6.25
Input
Output
Price reduction
GPT‑6 Sol vs. GPT‑5.6 Sol
$4 → $2
$20 → $10
50% cheaper
GPT‑6 Luna vs. GPT‑5.6 Luna
$0.20 → $0.10
$1.20 → $0.50
50% cheaper
Putting "ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC" into google search returns the result from Gemini with similar explanation, which it references to a Yahoo article about the Astra breakthrough and that's a result as of 3 hours ago.
Prompt
Decrypt encrypted with enigma `ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC`
Result
Raw Decryption (verbatim from the machine): BTTEUMANGABEDESMARSQWEGESXBEFINDEMIQINXROSENOWROSENOWXSOFORTFUNKANTWORTXWASCHBBSCH 2. Segmented German (with military telegraph conventions expanded): BTTE UM ANGABE DES MARSQWEGES . BEFINDE MIQ IN : ROSENOW ROSENOW . SOFORT FUNKANTWORT . WASCHBBSCH
(Standard German radio conventions: X represents punctuation/spacing; Q was frequently substituted for CH to prevent transmission ambiguities; BTTE was an enciphering operator typo for BITTE; WASCHBBSCH was an operator typo for the signature WASCHBUSCH).
3. Corrected German: „Bitte um Angabe des Marschweges. Befinde mich in Rosenow, Rosenow. Sofort Funkantwort. Waschbusch.“
4. English Translation: "Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio. Waschbusch."
Cryptographic Verification Re-encrypting the plaintext using the identical key settings reproduces the exact ciphertext byte-for-byte:
Enigma ( M raw ) = ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC Enigma(M raw
)=ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC 19:10
I ONLY know this because https://www.youtube.com/watch?v=JsBZOcqZerk, btw.
Even when the report literally says the LLM did it on its own?
Let's not over-correct in the direction of knowing better than the first party.
awesome! keep going
great work! keep going
"Carter Leffer only directed GPT–6 Astra to see if it could break any of the unbroken Enigma messages published on the Crypto Cellar Research web page."
"However, the most astonishing thing about this break is that the GPT–6 Astra did it entirely on its own."
> it suspected that the plaintext of Nr. 173, SIPVX, might be related to the plaintext of the unbroken MVUEH message
It makes sense that Nr. 172 and Nr. 173 might be related since they were sent at around the same time.
In Nr. 173, "ROSENOW ROSENOW" was also present.
It also makes sense that a longer crib would generally be more effective than a shorter one.
I am in Rosenow, Rosenow.
After trying many different approaches, GPT–6
Astra focused on using the repeated place name
ROSENOW ROSENOW as a crib.
(Is it possible that this is a misreported detail? It feels like a singular ROSENOW would be an equally effective crib)
Stubborn for a long time because the message used a completely different key from the rest of that day's traffic. Everyone assumed it shared the daily key. The original transcription had errors. The left rotor turned over at letter 72, which is rare and breaks standard crib attacks.
What is cool, if true, is that it was a 2 day collab between the Leffer and Astra. To me this shows the importance of human in the loop, was still all also showing how immensely power of llm tools. But I think it’s getting a bit silly how much anrticles ignores the driving force (the person) in breakthroughs like this.
Edit: Found it from here: https://mvueh-enigma-solved.carterl.chatgpt.site/
ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC BTTE UM ANGABE DES MARSQWEGES X BEFINDE MIQ IN X ROSENOW ROSENOW X SOFORT FUNKANTWORT X WASCHBBSCH
Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio. Waschbusch.Write a new prompt, evaluate, it now succeeds 99% of the time. Measure what changes between prompt #1 and prompt #2, understand what contributed to the performance jump.
Write a third prompt, this one succeeds 100% of the time. Increase the size of your evaluation set, find a 1/5000 error-class and a 1/10000 error-class, add some explicit code to correct for this cases.
Roll out to production, collecting usage metrics. You make some tweaks to your harness, your prompts. Eventually you have confidence that your system has fewer mistakes than 1 in 100k.
Now, multiply this iteration across all your different prompts and different ways that they might interact with one another.
This the state of software engineering in 2026.
Edit: clarified engineering to software engineering, which is more correct
Inside were internal docs, integration code, the Spaces app framework, memory records, container startup scripts, and documentation for an experimental ESP32-based home network bridge called Home Link. Codex CLI was also installed, though I found no evidence that Muse invokes it.
I didn’t demonstrate a sandbox escape or access to another user’s data. I reported the export to Meta’s bug bounty program, which marked it “Not Applicable.”
The post walks through the findings with screenshots.
-Pete
Using claude.ai and Opus, I asked "create a 3d animation from this" and pasted the animation SIML.[1] I just did that test again. There is significant improvement.
Opus 5.5 (high): https://claude.ai/artifact/5EgqfWcyVtLwJDQq6fsPUm
Opus 5 (high): https://claude.ai/public/artifacts/b37a9ee2-f5bc-4ff9-ae90-a...
[0] https://news.ycombinator.com/item?id=49526704
[1] https://news.ycombinator.com/item?id=49532609
Disclaimer: the skills and system prompt on claude.ai could have also improved, this is not a raw API call.