Blizzard is such a shell now, that I doubt we'd notice.
52,351 karma · joined June 24, 2014
Low-Thinking/non-Thinking absolutely has a place, but not in a tool like ChatGPT due to its very nature/designed purpose. Low-Thinking is useful in simple tasks/utilities where it is a straight 1:1 between the source and destination, like automated workflows.
ChatGPT Instant simply isn't worth using for the task they're assigning it. Medium Thinking is passable but High or better has a marked quality improvement/reduction in hallucinations.
Thinking isn't anything to do with logic/arithmetic/programming; it simply allows the LLM to spend longer deliberating/second-guessing itself, rather than looking for the shorter path to a supposed "answer." A LOT of mistakes get washed out in that second-guessing step (although of course mistakes can still occur YMMV). This lack of mistakes does make it better at logic/arithmetic/programming, but it also makes it better at everything else too.
I believe Google's Gemini gives you a handful of free Thinking credits a day, I'd give that a shot and I believe you'll see what I mean.
Ultimately to determine the underlying root-cause you'd still need to dig this same information out, and all they've done is moved the starting line behind several walls. In essence adding extra work, without solving this edge case/issue.
Regardless of if the information is in the BSOD, Event Log, or only via WinDbg, understanding the information relies on the expertise of the person reviewing/contextualizing it. They've gone out of their way to make contextualizing it harder.
For example, to determine if it is a direct failure or an associative failure (e.g. RAM failing causing different BSODs in unrelated modules), you want that context to be obvious. But without the module being in the Event Logs, you're now loading up half a dozen MiniDumps in WinDbg to find that same very key information - which people may miss or fail to do.
What I am saying is: If we believe that excuse (which I don't), they've done absolutely nothing to address it and just made that same problem worse with their childish games.
The actual interesting discussion, to me, is why Microsoft won't show WHO is dangling the handle open when the user tries to interact with a file via Windows' UI. To understand that, we have to look at a BSOD change Microsoft made in Windows 8:
In Windows 2K, XP, Vista, and 7 the BSOD would tell you exactly WHO was causing your BSOD (i.e. which module). Which was incredibly helpful, when you could see it was a e.g. Creative sound driver, or Nvidia graphics driver. Then in Windows 8/8.1 they went to the "sad face" simplified BSOD screen. From then on in order to see which module it originated in, you had to load the mini-dump into WinDbg (which almost no users would/could do).
What I am saying is: Microsoft went out of their way to shield their partners (OEMs/hardware vendors) from criticism with that BSOD UI change. So it seems unlikely they'd make a change to the "File Locked" UI that would essentially do the same thing: Open up their partners to criticism for their [bad] software (e.g. anti-virus/anti-malware/corporate compliance/etc).
Then tack on that Microsoft's own software may be some misbehaving software; and they'd essentially be telling on themselves. OneDrive in particular, I've seen in that list a lot (but I could write paragraphs on what a turd/abandonware OneDrive is).
> https://knowledge.workspace.google.com/admin/security/contex...
In particular "Allow access to devices using Chrome browser with security requirements" would present this message.
This is an unacceptable level of clickbait journalism. Nothing in the article's title is substantiated in the article's content, it doesn't break any of those three things, and the failures it does report are trivial (a dialog displays incorrectly, and a few devices have scattered reports of instability).
Microsoft has made a lot of mistakes in recent years, but this article isn't about them. We shouldn't invite in this level of clickbait even if it is popular to criticize Microsoft, because all it does is add noise to an otherwise very necessary discussion of MS's practices.
Just off the top of my head:
IFNA, FORMULATEXT, DAYS, CONCAT, IFS, SWITCH, XLOOKUP/XMATCH, FILTER, UNIQUE, LET, TEXTBEFORE/TEXTAFTER, LAMBDA, et al.
But my favorite improvement is the "don't intentionally corrupt CSVs" options found in Settings -> Data -> Automatic Data Conversion (hint: Disable everything). Only took them 30-years to add that. Absolutely absurd these are enabled by default still.
Excel is one of Microsoft's best pieces of software and one of the very few they haven't turned into slop YET. Still don't understand why we don't have local-only Python to replace VBA at all license levels (i.e. non-cloud).
So all of the legit providers will be required to collect ID, and anyone not willing to will be funnelled onto the sketchy providers; which I'm sure won't backfire at all...
Anonymity online is of course a double-edged sword, but we've seen the authorities, particularly but not exclusively, in the UK use intimidating tactics against those with unfavorable political views. Even when those views didn't break the law (e.g. no calls for violence).
If you also look at how nearly all the existing "verification" systems work, it is just a giant data drag-net, that is absolutely used to associate your real-ID with their advertising analytics. It isn't subtle. Which is why "big tech" (e.g. Meta, Google, Palantir) aren't far behind many proposals.
Neither of which ever goes away. These aren't short term costs, they're the costs of running their business, and it isn't profitable.
> The claim of the video is that they're losing money just serving current AI models.
Which is true. Every one is losing money, none are profitable. They're losing money serving current AI models.
> There's just no evidence of that.
Their own profit/loss statements are "evidence of that." According to these companies themselves, they're at a net loss every quarter. So it isn't clear what more "evidence" people need or expect.
It wasn't the correct way of handling the problem they were trying to address, but they definitely didn't hide it by any reasonable definition.
OpenAI, Anthropic, and Microsoft/Meta/Google are all at a net negative on AI (i.e. they're "demonstrably" losing money). So it is objectively true. If everyone is losing money, and nobody is profitable, then it is a demonstrable fact.
As far as I know, the only "AI" venture currently in the green is Nvidia, and they're selling shovels to gold miners.
If big AI does crash out, it would be an absolute gold-mine for local LLM. Cheap, efficient, Nvidia GPUs, and RAM that can run the best local models already available, will be a real boon.
PS - And as great as Qwen3.6-27B is, how large you can scale it (i.e. how big of a context/project) is mostly hardware constrained.
Who writes the class syllabus? Nobody from the publishers does, professors and or departments do that. Maybe based on advice from collage admin. But it is all in-house. Ultimately the college picks the books, they're the gatekeepers.
Calling it "inertia" feels very dismissive; and isn't close to an explanation of why somehow Higher Ed Professionals share no responsibility.
Specifically, the people making you waste money with bi-yearly re-releases, one-time-codes, or $150 textbooks, isn't actually the publishing houses. It is the gatekeepers at your very school: professors, department heads, and or the administration. Publishers are acting in an immoral way, but publishers by themselves have no power to force you into this abusive relationship. Your school is the one enforcing this, and yet few students file complaints at their school about the situation, protest, or otherwise make it an issue at THAT level. The level where they actually have leverage, and their complaints are more likely to be taken seriously.
Instead accepting the financial relationship forced upon them, and complaining that they wish publishing houses were less abusive. Publishers actually have little to no power themselves to force you into giving them money, your school does. So start complaining loudly and often at the school level if you want to see change. Every single year, every single class.
Without that, this just comes across like unconstructive commentary.
This moves the needle a little your proposals or the lack thereof don’t move it at all. So I’ll take this over nothing.
People are no longer commonly constrained by "model too dumb" limitations (in SOTA models). They're constrained by "model too expensive." So making the model ever so slightly smarter, while doubling the price, feels like a regression.
I actually think a Sonnet upgrade, while keeping the same price, would get more buzz. It addresses a wall a LOT of people, without unlimited budgets, are hitting (i.e. people feel forced to use Opus, which they cannot afford, because of Sonnet's limitations).
OpenAI recently retired Codex-5.3; which was very negatively received. Not because Codex-5.3 is superior to GPT 5.5, but because it was half the usage-cost while being "good enough." They made a better SOTA, but didn't realize that some of those customers are playing with Deepseek 4 Pro now instead of GPT 5.4/5.5 -- they were priced out.
Meaning the system prompt(s), harness, entry and exit points, and skills. So the product is still "Siri AI", because of all the stuff that takes it from a raw infrastructure concern upon up into a "product" is Apple's responsibility.
Google are "okay with that" because Apple pays them $1B a year, per press reports, to be.
Now, third party software, is always going to be all over the place. Stuff that was largely built on Win32 components works fine, but "modern" stylized applications rarely have strong support.
For example?
But that's a "Perfect is the enemy of good"-like argument. Wherein: Why even reduce an easy to exploit attack surface when there could be holes elsewhere?! Because, you know, it makes things much more secure even if imperfect.
Plus, to me, it is a culture issue. npm just doesn't take security seriously, so we don't see these improvements, and if there was additional test hardening later, I don't expect we'd see them in npm either. Since, they just don't care.
That's an INSANE default. pnpm, by contrast, allows you to essentially "opt-in" only specific packages that need this (e.g. four out of thirty, in one of our projects). Then tacks on tons of other security settings, like minimum age, no trust downgrade, etc etc.
All attackers can attack packages by updating how a package functions; but npm is particularly problematic as it runs non-sandbox scripts as the calling user. Putting not just your project at risk, but your entire machine/network.
And this stuff has been known about for YEARS, they've taken no action.
We'll see how it winds up, but we could see models get licensed over half a dozen+ compute vendors, and then you pick your price/offering/features favorite.
While I'm also a huge fan of local LLMs and believe they will be key in the future; I think the claim of "just as good" is hyperbole. They're productively useful tools though, and something worth exploration.
Right, but you still need to connect that virtual machine to their service/servers in order to actually accomplish anything. This change doesn't move the needle of where you were before.
> Went out of my way to patch the ELF itself because the prompts are hard coded.
Why even pay for Claude Code at that point? CC is MORE expensive than many competitors, but it is popular because they take care of all the hard parts, creating a very high quality "turn key" product. If you're putting in all this effort, may have well just use OpenCode and one of many API vendors.
> They're literally injecting strings from the network into the system prompt. If it's not prompt injection, then I have no idea what it is.
I agree you have no idea what prompt injection is. Here is the Wikipedia Article's first line (which I agree with, as a definition):
> Prompt injection is a cybersecurity exploit and an attack vector in which innocuous-looking inputs (i.e. prompts) are designed to cause unintended behavior in machine learning models, particularly large language models (LLMs).
Anthropic are sending down a system prompt to their proprietary software from their proprietary service. It isn't an exploit, isn't an attack vector, and isn't unintended or unexpected.
> I can only tell you what I'm doing. Here you go: https://github.com/matheusmoreira/.files/blob/master/%7E/.lo...
Those seem like pretty reasonable changes to the prompt. Why is altering the system prompt more effective than instructions after?
Can you clarify why? You decided to install Anthropic's software (Claude Code extension and or CLI), and then utilize their service which you're paying them money for (and have a contractual relationship with). The software itself manages tool-usage safety/sandboxing, so you're kind of trusting Anthropic a LOT already.
Why does moving the system prompt from within their proprietary software, to their proprietary backend, matter at all for Claude Code users? It doesn't feel like "hack the Claude Code binary to alter how it works" is a common and or supported use-case. Most people pay Anthropic so that Anthropic takes care of that stuff, and lets them get on with their work.
Also; I'm also not sure if this meets the common definition of "prompt injection." The vendor you're connected to is sending a system prompt to work with their own model/service. Where the system prompt is stored is immaterial.
PS - My gut tells me there is something else going on, leading people to hack the Claude Code prompt/binary. And that the "something else" isn't supported by Anthropic.
There isn't cryptography true-believers, it is just mathematical proofs with sometimes good or bad code wrapping it.