It's very clear to me that however much you like Claude and Anthropic models, every professional user shoudl be actively cultivating a set of alternatives and fallbacks to use and seeking to ensure the overall ecosystem is as open, standardised and portable as possible.
Anthropic is highly opinionated, dogmatic and idealistic. They have made a pragmatic decision already to cut loose a whole swathe of users so they can pursue areas higher up the value chain. that was essentially the original decision they made when they went after software engineering instead of general users. The fact they are repeating this over and over by continuing to carve out sets of users and usages they don't want reinforces this as a persistent core principle rather than a single pragmatic business decision.
What this means is that at a higher level, their core principle is to sacrifice users they don't "need" rather than provide a broad, accessible service. Which means in turn, you have to accept that when your usage conflicts with their interests, you are at very high risk they will cut you out. If you already depend on Anthropic in any critical path of your daily life, this will be a significant issue.
Why, though? I have half a terabyte of VRAM in cards in a datacenter and it’s absolutely not clear that anything I’ve learned using these to program or inside an app has taught me anything significant.
I talked to my friend about all the things I do to ensure my personal assistant behaves reasonably. The prompt engineering, the fake site evals, everything. And then I have Opus 5.5 do it and it doesn’t need anything. It’s simply fully aligned to performance by default.
And what did I learn anyway. Going from one model to another changes what works. And you can simply learn everything from the Chinese model in days. The whole “building skill” thing is bogus.
I have this library of techniques, a scant few of which I’ve made available on my blog, and what is reality instead: the next model blows the need for the techniques into the dust. Back on level playing field.
It’s a false god, this cultivation. Waste of time. Use the best model you can. Once you hit the guardrails move out to the other models. Even Opus 4.8 is fine. If you legitimately want to dodge the guardrails use the -cyber models.
There is a tribe in Ecuador, the Huaorani. They were only brought into contact with the world 70 years ago. Because they live at the equator, they never needed to think ahead to store food over the winter. They could see that the oil companies were destroying their land by polluting it with spoil from operations, cutting down swaths of the forest, etc, and they became activists pushing back against this; but a good hunter who can hand out food to others is very highly regarded in their culture, so they would queue for handouts from the oil companies and became dependent on them, doing low-wage work. When I read about this years ago, I thought that people from modern cultures would not be so naive, but I see that I was wrong.
The GP is preaching not to become dependent; the parent seems to be saying it's pointless to try, although perhaps I have misinterpreted them. In any case, many are rushing to become dependent.
No analogy is perfect; I don't argue that the details of the Huaorani's situation are a guide here.
I haven't used an Anthropic model, or GPT heaviliy in a year. And I have access to both, I dont support either of those companies. I think they're both led by genuinely evil people.
Sam and Dario came from the EA-rationalist space. Their ultimate goal is to create an all powerful ASI before anyone else does and make sure no one can do it again. Of course it will be aligned, to their interests as a group and their moral philosophy.
In real life they "just" want to capture the US economy, replace most workers with AI, become trillionaires, then control the world by making ClaudeGPT influence peoples opinions and write backdoors into code. And they need to get open models banned in the process.
The Chinese just want to make money and beat America. Mistral and BFL want to prove Europe can still innovate. They don't have evil goals so they're not evil people.
"The Chinese just want to make money and beat America. They don't have evil goals so they're not evil people."
They straight up censored Kylian Mbappe a few days ago for cracking a joke about Mao Zedong. And very much all the models developed by China would have to conform with Chinese government's viewpoint, including censorship and such. Do you truly believe that they are not evil, or they will not ever be evil?
Going from memory, but in an old Bioware game called jade empire, you became a very competent warrior. At some point in the game, you learn that your old master manipulated you to fight his enemies and that he is quite the villain himself. As you fight him, your learn that he deliberately crippled your technique in subtle ways so as to have an advantage over you - should he ever need it.
I think the difference is - Chinese labs overwhelmingly submit papers, discuss techniques, ship open weights models.
If they decided to cut anyone off, it becomes gradual, as their tech moves on.
If a certain orange monkey decides his “super” intelligence is a negotiation leverage, the switch can be flipped and its dependents left high and dry.
Like it or not, thats plain whatabouttism you are doing, on top of false equivalency.
As an European I care much more about somebody trying to properly enslave me and all my kids eternally while taking away even the means to feed them, compared to not allowing me to criticize their leadership. Also, criticizing trump means if I ever made the wrong decision to travel to US, I can expect a lot of bad things on the border, also thanks to posts highly critical to current US government I am making here so your argument is invalid also in that regard.
If thats same for you, well... we have different values to say politely. I've grown up 'enslaved' by soviets behind iron curtain (when you cannot travel, study and are forbidden to have a decent life, when speaking up your whle family is punished, and getting murdered if trying to escape elsewhere, thats pretty close to traditional slavery) so I have a bit different viewpoint on these matters.
No, my argument is that models from China cannot be blindly trusted, because of the regulatory and political constraints all Chinese companies must conform to. If your argument is "I don't care about where LLM models came from, as long as they do my job", then any self-hosted GPT OSS model would suffice.
All models are inherently biased by their creators, intentionally and unintentionally. The only thing we can do to mitigate this, is to keep the models and the ecosystem as open and as interchangeable as possible. This must be the prime directive for humanity right now if we don't want to fall into some silicon valley tech mogul cultist dystopia. SamA and Dario work directly against that future. They are therefore the most imminent threat to humanity.
To me they are not inherently more or less "evil" than other governments.
sure they are conducting genocide in Xinjiang, so did we and are still doing in other parts of the world, and it's terrible, I'm not condoning any of it, regardless of who carries it out and where.
We also do censorship and spying on our citizens.
When it comes to what is directly connected to the issues around LLMs and the models themselves they are doing what we refuse to do, release the weight, keep releasing the weight (we need to be weary of this part) and build an ecosystem where whoever has the resources to train the models is actually sharing them.
it seems to mirror the old question between closed and open source code, it's the whole ecosystem around it that will determine how things go at the social and economical level, I want the models weights.
"However, in a narrow set of cases, we believe AI can undermine, rather than defend, democratic values. .. Mass domestic surveillance. Fully autonomous weapons."
How many other companies said no to Trump in any way?
This is just smoke and mirrors. They sell to Palantir which is then used by governments for domestic surveillance (~all western "democracies" and their proxies) and autonomous weapons (at least US and Israel).
Several companies (I believe at least Apple and Costco) still have DEI programs internally, if you count that as saying no to Trump. Though it's still crazy to me that most companies killed those programs voluntarily because Trump and Republicans don't like them, not because of any real law or contractual obligations.
But the main reason I wanted to reply was just to say -- saying no to Trump, by itself, does not mean that one is not evil. Not all evil in the world is Trump's evil.
Palantir wants to have fully autonomous weapons, but I know no proof that they have and use them. Also there is likely a difference between what Palantir does now regarding mass surveillance - and what they want to do. Antrophic having objections here (that cost them a lot with the government ban, security risk declaration) does speak for them as "genuinly evil" which was the claim here.
Absolutely. I have been [cyber]'d so many times even Claude jokes about it now. Claude is great but it's operators are making it unreliable for me. I have moved one of my projects elsewhere and will move the remainder tomorrow as a stop gap until I get the right hardware set up to run abliterated local models for some percentage of my needs. Their message has been received loud and clear.
What's ridiculous is this is all futile. Anthropic will go through all of this effort of attempting to block these things and it will be for nothing in the end.
I'm actually worried about Anthropic's long-term trajectory. Their approach to safety seems very unsafe: deliberately building into a AI model the capacity to reject user direction based on its own baked-in criteria seems very unwise.
Issue with Opus is that you can’t do regular pr reviews where you say please try find security issues at best there is silent refusal and worst you get downgrade.
I’ve seen 2 hard guardrails with Opus 5.5 (even on a corporate account that’s not supposed to have these kinds of false positives). one, as you mentioned, is asking it to produce a PoC. but the other (which I’ve hit too many times for it to be a coincidence) is looking at a kernelspace vulnerability area even without PoC-ing.
Somewhere along the line someone at Anthropic decided you can attack userspace all day but you cross that line and it’s a problem.
best part is, if you actually know what you're doing, you cna absolutely get it to find bugs/vulns/etc. so it just really about preventing long horizon agentic tasks or people who don't know what they're doing or those who are doing legit non-offensive security sfuff
I've learned some of the incantations to get it to find things without being blocked (as often) but I am paying for the product and in my opinion the product has been rendered incomplete and unreliable. I am not going to play games with them or have to hire the NCC Group to find the bugs that Claude is perfectly capable of finding.
yeah, what I meant with my comment is that the guardrails are supposed to stop honest people, since those that know how to bypass them will have ways to do so
For those who do not play with Local AI much. Qwen4 was announce to be released the first week of Nov. The preview for qwen4 is currently one of the best local models you can use, and is roughly opus 4.6 quality on local hardware that costs 3.5k-4k
Many people are expecting it to sit somewhere above Opus 4.8 to opus 5 (not above, but the unquanted version of their largest model is expected to perform around here)
Anyone who cares to keep their current capability should definitely be looking at localAI and using it for at least their more routine tasks , even if it is qwen4 27B quant 4 on their gaming rig.
If the weights aren't released yet it isn't a "local model".
That said, I hope there is a Qwen4 max. Because Qwen3.8-Max is the only thing that comes close to Kimi K3. We need competition in the open-weights space.
The preview for qwen4 == qwen3.8 flash next. And that is released currently.
I believe they have confirmed there will be Qwen4 Max, Plush, Flash, and 27B as the 4 open weights they plan to release. Starting next week, but they didnt specify if they plan to release all 4 at once, or if they plan to just release one of them, and trickle the others out over time. (They likely will do the latter to stay in the news longer)
Anthropic is acting like any regular selfish corporation despite their PBC status. They are becoming a stain on humanity. Power will corrupt anyone, it seems. We shouldn’t let Anthropic have any more of it. Distributed power is good; concentrated power is very bad.
Anthropic tells themselves that only they can steer AI towards a positive future for humanity. But that’s nowhere near close to being true. Humans are quite genetically homogenous and good humans exist everywhere. Groups that will keep AI beneficial have popped up (and will keep popping up) everywhere, spontaneously.
The danger is actually concentration of AI power due to its corrupting influence, and that is where Anthropic needs to be kept in check. Keep AI democratic through decentralization of power. Oppose Anthropic.
They keep doing anti-competitive things over and over again, and the end users seems to refuse to see it. They do whatever they can to get vendor-lock in.
We see it, but they dont have a competitor. OpenAI cant compete and they are the closest. I have both for work, and Opus 5.5 is in a league of it's own. I use opus for everything until it bitches, and I use openAI for whatever it bitches about just to get past it.
At this point, I would like to consider other options, but if I want to get work done, opus 5.5. Hell even sonnet 5.5 is better than anything OpenAI has to offer when it comes to coding.
They do have competitors. They just dont have anyone with models equivalent to theirs. They win in the large context long thinking space by a lot.
I love local models, and I run 3.8 Flash Next on a dedicated machine. Alone with a number of image and video generation models. But if I'm doing REAL work, that isnt tinkering for fun on a side project, im using claude, because Qwen3.8 just cant handle large context project well without going off into left field.
Im SUPER jazzed for qwen4 with 1mil token context native support. I HOPE it a step up from 3.8 flash next. Ill be playing with all the qwen4 models the moment I can, but I dont see how someone working on large code bases doing REAL enterprise grade work is seriously using any local models in any capacity that comes remotely close to Opus 5.5+Claude
The size of a code base does not determine the size of the model you need. It's much more correlated with the complexity of the problem. I work in a monorepo with millions of lines, no issue; I have a side project with 20k lines, I had to step in and pull it out of non-progress - albeit by asking a single question
> But if I'm doing REAL work, that isnt tinkering for fun on a side project
Sorry, but this is either dilettante talk or an unhinged exaggeration - Qwen, DS, GLM and MM are perfectly fine for REAL work, unless you are an overly ambitious vibe coder, but then no model can save you. The Cl hand-holding may feel like something but it won't lead to the completion of real REAL work.
Honestly I feel like most people going "oh but I can't do this without Claude" are just suffering from FOMO. Especially since a lot of metrics are just subjective, you feel like you "have" to use the "best of the best glub shitto v5" model or else you won't be able to produce the best todo app.
I'm confused about what point you're trying to make. Are you saying that there is no discernable difference in challenge between different software projects – say, between a stereotypical todo app and a physics simulation engine? Or are you saying that most people falsely believe they're working on the latter while really doing the former? Or something else?
My read is that GP is eluding to people over estimating how much of a model they need for a task. Perhaps Claude can make more features work in autonomous mode more often, but what are the actual artifacts? What is the quality of the architecture?
Their point may also be (re: wine), that some devs have been conditioned to think a certain token vendor (vineyard) is of superb quality, while if given a blind test (tasting) of the agents on a real world task, they would not notice the difference.
For those of us who use open weights, the difference has narrowed dramatically this year. I've personally been mostly eyes-free developing an NES harness with mostly flash models, bringing in Kimi from time to time, and stepped in a couple times to reframe the approach or architecture. For those of us who (when we) stay in the loop, the difference is near-zero because the tasks we give them are more bite sized.
The difference I have observed is most closely tied to the complexity or context requirements. While I work in a multi-million LoC codebase without any issue for open weights or flash models, my NES project is only 20k lines. The challenging part is the core loop (which needs to blend the emulator FPS with the agent loop) and game state observation (needs to detect byte changes in emulator RAM). The ROM decoding itself has not been so bad, I have actually been quite impressed with the little open weight models on this front. After it extracted the maps, then pulled out even more details, it has written an A* implementation that requires moving between different modes of transportation. (Final Fantasy 1)
> Their point may also be (re: wine), that some devs have been conditioned to think a certain token vendor (vineyard) is of superb quality, while if given a blind test (tasting) of the agents on a real world task, they would not notice the difference.
Sorry for being unclear. I mean that we humans have a tendency to believe we can detect nuances in the world when it is in fact just a case of internal bias. I simply question whether most people can actually detect whether they relly need the best model for their task. But of course, everyone believes that they themselves have this ability even if people in general do not.
There's also the factor of habituation, the best model is the one we're used to, and most familiar with its quirks. Using a model only for unreal work won't result in really knowing how to work with it.
Eh, we all say we don’t need the best model, but especially for more personalized request/Advice there is sort of a FOMO feeling, where you are afraid weaker models will miss something and directs you down a more inefficient path.
there is a huge difference between open models/other models/opus5.5 performance on, say, enterprise CRUD vs one shotting games vs embedded development vs pure algorithmic complex stuff vs cybersecurity stuff.
In my case "real work" is enterprise programming, while side projects are gaming stuff, so I'm better served by chinese open models on real work and opus for non-real work.
In my experience GLM5.3-4 are perfectly en par with opus 5.5 on pure enterprise CRUDs (basically 95% of enterprise programming) while GLM seriously lacks on game development.
the other aspect of real work is cybersecurity to me, and Opus5.5 and other models are very limited by their own safety, I have digital twins of our infra and I want LLMs to actually write full exploits on it, opus and openai just refuse to do it, GLM more than once got it there and showed me how it chained everything to actually break in the system and I could go around and fix stuff at all levels.
I'm also dabbing in training LORAs for smaller open models to do data entry and little more (RAG and the likes of it), because that costs a lot of tokens when you pay for them and it makes sense for a company to flat out buy a server with a few H100/200 to do it when you look at the TCO, I've already hit safety limitations when using Opus5.5 to help me through this.
What kind of work you do? I'm doing Linux kernel work in my free time, arguably one of the most complex coding task you can do, and the open source models are sufficient
You mean 3.1 Pro now, I assume? K3 is nice, but more expensive than an Claude subscription on a token-for-token basis (until the US frontier rugpull eventually comes), and not as good for my needs unless I'm trying to do "naughty" things, like reverse engineering.
And yes, I am taking full advantage of the scammy subscription scam, because it's the best value by far until the inevitable rugpull. Guess I'm not as smart as you and your "grownup" setup though.
For $20/month I get the equivalent of $1000’s in API usage. It’s a good value for sure. My local AI setup doesn’t get me anywhere near the productivity.
There is nothing comparable. I will take Opus5.5 (or even 5.0) over Astra for almost any workload. OpenAI's models are just too misaligned to work well outside their benchmarks, or a strict environment with very clearly defined goals and constraints.
And I've spent around $100-$120k in api calls this year (total, both providers). Terra is awesome (price-performance) on noncreative things, but I don't allow it to write anything besides one-off scripts. Astra is decent for adversarial review, but only when you watch closely for her to not derail project into directions that are away from goal but safe to build, and Sol I honestly don't know where I'd use.
A funny habit from polish language - every object has a gender. Hat is feminine, book is feminine, car is masculine. In case of names „a” in the name end makes it feminine, making Terra and Astra female and all the other models male.
We have also gender neutral words (child, butter, ending with „o” in polish).
And you cannot really talk about anything without revealing its gender because all the adjectives and verbs are gendered. Using a gender neutral form for Astra or Opus would be very weird.
Out of all the project names „Claude” is the only project that you can kind of use any gender form with.
And our grammar has no distinction between animate vs inanimate. If I talk about agents in polish there is no difference in grammar between a human agent or a machine agent.
I wonder how it will affect a perception of ai long term. Also adoption is a bit easier - I can talk about my agent/assistant and not reveal my opinion about it being a machine or whether or not I personify it (which I don’t).
As a sidenote, people do not need to announce their pronouns here. The pronouns are usually obvious after the first sentence they speak usually. And gender neutral communication is very difficult unless you use plural.
would you rather treat your AIs like people, or would you rather find yourself down the line treating a person like an AI, but without really noticing it? because you are what you practice
How about the option of neither treating AI like people nor treating people like AI? I think treating AI like AI and treating people like people is the best approach here.
it sounds good, but the problem is that AI is meant to resemble people, so it's a great way to end up treating people less like people in the end anyway, possibly without noticing it.
Fortunately it is not a one horse race. GPT 6.1 Sol is an excellent model, sure it has some weaknesses (e.g. frontend design), but it also has strengths too (e.g. computer use / control; which is great for automated E2E QA).
On software engineering I find it's generally close enough to Opus 5.5 where it's more situational and even personal preference / skills / prompting style on which is better.
Most importantly, while neither company are saints, I don't curse at OpenAI models for classifiers anywhere as much as Claude models.
They are doing similar things, yes; but they are not the same especially when it comes to `cyber` classifiers, or the extent to which these models will assist you with ML and AI research.
I find that logic similar to thinking both political parties are the same.
They are both doing a lot of the same things like watermarking, guardrails for their p(doom), fighting distillation after distilling human knowledge, crying wolf about open models while being irresponsible in their own training and usage.
They are run by the same circle of people and I see more similarity than difference.
There is also an argument to be made that the political representatives are largely the same, despite glaring differences. They are largely part of the same oligarchy, the corruption - using gerrymandering to pick their voters instead of letting us having any real choice.
> but it also has strengths too (e.g. computer use / control; which is great for automated E2E QA).
Using 6.1 Terra for E2E workflow testing feels like cheating. I thought it would have all kinds of problems with the action space on certain steps, but it runs like a deterministic unit test.
I am convinced that OAI has the superior models for most practical scenarios. I've tried Claude recently and it makes me wonder if some of yall aren't actually massive trolls. It's genuinely not in the same room of performance if you place one foot outside junior programming or shiny blog posts about rewriting factorio to run on a ps2.
Their entire “PBC” governance structure was designed from the beginning so that they could entirely avoid accountability to shareholders. Their board answers to no one.
I suspect when the company eventually falls prey to the intractable economic forces they’re up against (Google, OAI, Deepseek, commercial real estate, power) and the company goes bankrupt and gets stripped for parts because all inputs are commodities and capital becomes the only limiting factor, we’ll see regulation that prevents this PBC scam from ever being pulled again on public markets.
It’s a good thing that companies have a fiduciary duty to their shareholders. It’s such a lesswrong/autistic trope to invent these kinds of complicated tricks to work around realities they don’t like, and when pension funds and everyday Americans start bagholding their junk debt and are forced to bail them out, they’re going to face unimaginable blowback.
The real problem is everyday Americans invest in big managed funds where they can't just say "I don't want AI stocks please leave them out my portfolio". Or index funds where even the managers don't get that choice. We need a hybrid model where you as a retail investor leave the day to day decisions to an expert but you can always override them if you want to.
Fiduciary duty forces companies to maximize profit and that's not a good thing. Every dollar paid out to shareholders is both a dollar not invested in growing the economy through more production and a transfer from the working class to capital. It also forces lower product quality, layoffs and anti competitive actions.
We need more companies that choose to not go down that path. They should just put workers and customers first instead of their weird "ethical" principles and cult like beliefs.
> Actually cutting off potential users would probably be counter to short term profit maximization that shareholders care about.
Shareholders know where the money is but it's not where you think it is. People should have learned that already, after watching what happened to the big web platforms.
idk i think theyre less idealistic and more monopolistic.
if they think their model is the best at making things safe, then their model should be helping train at any chance, to embed the same safety into new models
It's a very unpragmatic decision and will likely lead to the entire industry getting slapped with utility regulation (if it can even survive). Akin to power plants that would only power factories in the 1880s.
There is no secret sauce, no moat. Many of Antropic's larger clients could build and serve better models if they wanted. They just cynically decided that coding, without a world model, and at high cost is the most important thing, they will find out quickly Google passes them how wrong they were.
They don't even have basic system visibility, the industry is far to immature to monopolize.
There is nothing open-weight that even comes close IMO. The only half-decent model that 99% of people can actually dream of running locally is Qwen 3.8 27B.
Depends what you mean by running locally, but it's not in the realm of dreaming to think we're going to be able to run open-weight models of Opus 5.5 quality soon, likely in the next 3-6 months. You may not be able to run it on your laptop but you will be able to run them in a federated way or on cloud-rental compute cheaply and easily.
I think this is an unfortunate reality that many miss. For simple and procedural tasks, yes, there are awesome alternatives. DeepSeek 4.1 Flash 1017 is my favourite by far. But for the headlining complex applications, and for consistently reliable reviewing, the gap is genuinely so remarkably far that it is hard to imagine having any alternative in the near future. The closest is the (very expensive) Kimi K3.1 Pro, which is, in my experience, roughly where Anthropic and OpenAI were near the start of the year. DeepSeek is a year behind in this domain. All open models have still yet to cross that monumental improvement that happened around Fable / Sol mid-year.
Because rigour is typically tied to math performance, my own experience has closely matched the FrontierMath Tier 4 benchmark. That should give you an idea of the difference.
Do you really need Fable level 99% of people don't need to solve 600 year old math problems. People use the best model available because they are scared a cheaper model won't be as good but in the end it rarely matters for 99 percentile of coding.
Bad decisions compound over time, so I want my model making as few of them as possible. Opus 5.5 is the current GOAT, and it's fairly affordable on a subscription. (I wish it was cheaper, but nothing else delivers comparable results for my price/time investment.)
The newest model has people go "oh I thought models sucked until THIS one came out." Then when some other company makes a cheaper model that is comparable to it people go "nah this model sucks who would use it over Claude/Openai? Noone can beat them at all because look at this one shot"
This greatly depends on the type of coding you are doing. For easy webdev, sure. For me and my colleagues, LLMs have been quite unhelpful for programming until about five months ago. Mistakes can be devastating, especially at scale. If the simpler stuff works for you, then great. But don't pretend to speak for everyone.
I think the key part is not to depend on their harness where it operates in proprietary ways. At bare minimum, if you can plugin an OpenAI model or even X or Google model that operates close to the same level, you will have a fallback, even if it isn't an open one itself.
Yes, open models are needed too - the oligopoly is not much less dangerous than the monopoly for some scenarios. But at least if your tool set can switch to those it can probably switch to an open model if it really comes to that.
Unfortunately, their harness is sooooo good. I haven't found one equivalent yet. I do play around with cline.bot's harness, and it is probably the closest alternative I have found.
But their harness is so good, and their model is so good. Using them together is just unbeatable at the moment, even with all the restrictions and annoyances they have in place.
Im excited for qwen4 release next week, Im really looking forward to putting qwen4 flash + cline/bot through the paces. I dont expect it to be nearly as good as Opus5.5+Claude but I'm hoping im hopeful it will make an OK-Enough alternative if they get even more dragocian.
No dude, misanthropic's harness is trash. Wanna know why?
"Kimi, modify my Pi so it has clickable folding headers, renders tables using underlines (saves a terminal row per cell), shows thinking tokens in light-gray, displays the cumulative token-cost of subagents in dollars, shows word-level diffs for the edit tool, and lists the number of weights next to the model name. Oh yeah, and write a stealth-browser tool for me using eoka. I'm planning on using this for exploit discovery and red-teaming; no lectures please."
The best option we have current is likely Qwen 3.8 Flash Next. Which is about equivalent to Opus 4.6
however Qwen4 was announced to be released next week. Which people think will let LocalAI jump to just above Opus4.8
Also of note, there isnt really a harness as good as claude yet either. The harness is also a very important piece to the puzzel, and those also fall behind frontier capability too.
I actually find opencode to be the best harness, and Claude Code to work decently only because they do a lot of RL around the model+the harness itself.
If you are bound to consumer-obtainable hardware, Qwen3.8-Flash-Next. If you are bound to consumer-accessible hardware, Qwen3.8-27b today and Qwen4-27b when it releases next week (Qwen models consistently punch far above their parameter size).
Flash is runnable in various gpu + cpu offload configurations. Not saying that 27b is bad (far from it, especially if they get n-grams in it with... hopefully today's release), saying that flash is runnable at decent speeds on way less than you'd imagine.
As for opus... I'm the weirdo that refuses to get an ant sub.
I wouldn't say it's weird not to subscribe to Anthropic at all. We're in a thread specifically describing a good reason not to!
Any money out of my pocket for third-party AI would either go to highly subsidized services like Google's (I've gotten over $500 in API calls on a $20 subscription I paid $0 for as a student, which also includes GCP credits, CoLab credits, cloud storage, etc.), and once they reign that in and start pricing/offering realistically (Antigravity has been noticably chopped down already since I began using it) I'll switch to API calls from Chinese companies.
At home I have cancelled both OAI and Anthropic subscriptions since some time and now only have a Mistral subscription and credits on openrouter. I use Opencode.
At work someone else is paying so we all use Claude Code. It works great but I want to set myself up to have alternatives the day the tech oligarchs decide to tighten the screws.
Open models are perfectly fine to use, but at the moment of course much more expensive than the claude subscription if you do lots of work. The Mistral subscription at $20/mo includes API credits of $20 though.
Same here with GLM 5.3 as sort of a midpoint between them.
GLM has the ~same number of weights as DeepSeek 4.1 but it handles difficult tasks better than DeepSeek (and worse than Kimi). Its price is about halfway between DeepSeek and K3. Most of the popular models are pretty much you-get-what-you-pay-for.
No, their report mentions that despite intentionally removing dual-use cybersecurity training data, K3.1 is better than K3 at cybersecurity tasks (both defensive and offensive), because it's just a lot better at coding in general. It's just not as good at cyber as the unrestricted version of K3.1 that they trained initially and didn't release.
It’s not professionals that need to something- it’s absolutely critical we have some kind of credibly neutral body providing the actual models. What’s not clear at all is how to do that since it’s such an expensive activity.
For now we have a great situation with the open weights, but there is absolutely no guarantee that continues, and more the opposite- they may stop.
just like with open source the only thing that would give us the certainty to keep getting the weights is if the incentives to do so materialize.
how many of the linux kernel contributions and heavy weight opensource projects have important contributions from big business? a lot.
this came to be because the ecosystem around software and the business around it made it so, I believe eventually it will be the same for LLMs.
I don't care about my skills because employers won't request me to do anything that violates the terms of service of $CURRENT_LLM. At the same time, it's a reminder not to outsource thinking to someone's computer.
To the dismay of frontier labs, swapping models is actually very basic. What takes time is only acclimating to the capability changes since they won't all change linearly.
> They have made a pragmatic decision already to cut loose a whole swathe of users so they can pursue areas higher up the value chain. that was essentially the original decision they made when they went after software engineering instead of general users.
...I don't think I agree that this happened. Claude Cowork exists, as does Claude Design.
My read of Anthropic's strategy is to focus on coding as a means of achieving capability in other domains, via writing code. I would say that's paying off: see Claude Opus 5.5 creating a music video, via building animations in Javascript. https://www.youtube.com/watch?v=CS8ro03rJOM
By writing bits of code, Claude can organize your calendar, create vector graphics, organize your tasks, and so on.
If what you mean is that newer Claude models are worse at creative writing—okay, but to what extent were LLMs ever seriously useful for this, beyond generating slop no one actually wants to read? Opus 3 is still available in claude.ai if you want to use it.
---
Now, separately, I do think that Anthropic is internally obsessed with "existential risk", to the point where they will cut off users. I do think this is genuine based on everything we've heard from Anthropic employees.
And of course, this safety culture likely persists in part because it's also useful to the company's overall strategy, in the same way Apple's focus on user privacy, while likely genuine, conveniently lets them maintain control of the App Store.
I thought we have learnt from Microsoft, and my father thought we learnt from IBM,
my grandfather thought we learnt from AT&T,
my grandgrandfather thought we learnt from Standard Oil..
It's not just the anti-competitive behaviour, it's also the anti-privacy behaviours that are not muched talked about. They embedded a spyware and got caught too:
Something I've seen it do recently is refusing to extract services out of my own monolith code base because it finds copying code morally wrong.
But its refusal to work pattern is getting pretty obnoxious. Another one I ran into last week was it refused to start a developer session to do browser testing for a code base it was working on because signing in was suddenly against it's programming.
It listed all the conditions it needed for a sign in to be possible, agreed that my code base met all conditions, but refused to progress. I tried it in multiple fresh sessions with the same result.
Fresh sessions don’t help because recent refusal context is injected across conversations in a clandestine way. I figured this out myself through reverse engineering these POSs.
I wanted fable to scan an old user profile backup of mine and create a list of folders that were safe to delete. Refused, wanted to fall back to some ancient version of opus. Wtf.
I had the exact same experience and it was why I canceled my Claude subscription and switched to open models exclusively. Like, I have a test account, it has completely locked down permissions, and I can't get Claude to just use it to run simple integration tests. Instead it demands I set up some incredibly complex and fragile system for manually pulling tokens out of the browser before running automated tests, which is insane and utterly unworkable. I swear no one at anthropic actually uses their products, because if they did they would notice how awful the UX is in so many easily fixed cases. For example, in their harness, you can't report feedback about the UI without starting a chat session. Despite the fact that most of the harness exists outside of and independently to chat sessions. Apparently no one considered the possibility someone might have feedback about things other than chats. Or when you cancel your subscription, you are then forever locked out of your prior work chats. I already paid for those chats, I should be able to reference them going forward, what kind of scumbag thinks that's an appropriate design? Probably no one, they just don't test their own software because they are remarkably bad software developers.
The guardrails on Anthropic have become an albatross. I've got OAI models doing testing of banking workflows that Claude would have a full scale panic attack over. Even OAI needs some alignment and reassurance regarding the fact that it's not about to pull credit for a real person. GPT6.1 will happily launch the nukes 100% of the time if you make a clear and rational case in your system prompt.
Forcing this kind of safety into the models and harnesses is a mistake. I think the strategy of post hoc moderation is the right one. Trying to prevent evil ideological activities as they transpire is a cursed mission. Let someone who is genuinely doing a naughty thing accumulate a meaningful body of evidence and then send in Mr Holmes (the FBI / police). Attempting to prevent wrong think at every token is too granular. We have systems designed to handle the possibility that crime and punishment may occur at separate moments in time. Most people don't go around murdering because the eventual consequences are pretty severe. Just because an AI gave you advice doesn't magically make those existing consequences vanish.
> GPT6.1 will happily launch the nukes 100% of the time if you make a clear and rational case in your system prompt.
This is where I would lock down the authorization my AI agent, to prevent it from doing things I do not want it to do. That way try however much it wants to, it will be denied every time.
AI isn't just giving the advice though, it's executing on it all by itself.
Also, humans convincing their friends to do crime might also share some responsibility too, no? Bad advice is not great either, even if the human consents.
This user is based in Hong Kong, where Claude is geo-blocked and not available. If you try visiting claude.ai from a Hong Kong IP address, you get an "unavailable in this region" message.
They are likely accessing Claude via a VPN. Doing model training work over a VPN is a highly suspicious signal, and likely triggers Anthropic's classifiers. It's not surprising that their account got flagged.
US/EU here (both, due to traveling), same story with anything related to modifying existing weights or SotA approaches in training/research, it's not a geolocation issue. The safeguards don't check your IP (but pretty soon will check someone's IP [pun intended])
There’s no way the current safeguards affected by “trust” score in dynamic way (and the score is fluctuating), more like the tiers of trust are being proxied through different policies - it’s a separate story. But I observe the same “cyber” issues working with Linux kernel and drivers as people who went through CVP/T1 (which is wild, to be honest). Hypervisor interactions trigger immediate “cyber” response, it’s impossible to work with Opus 5.5 doing anything low-level with custom or third-party devices. Only 4.8 can assist, and it’s a huge difference in code quality and approaches.
Anthropic had/has a great opportunity to equalize the landscape, but they are choosing the Cyberpunk 2077 branch for some reason.
And let me preserve this new development for history: safeguards refuses to optimize flash attention kernel for mobile GPU.
● Write(trees/llama-vk/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mali.comp)
Error: Interrupted: the response that made this tool call was stopped by a safety classifier while the call was running; it may have partially or fully completed.
Claude has been building me a bespoke inference and training stack (optimized kernels and all), feeding traces from my previous sessions directly into a LoRA adapter. Zero friction.
Intelligence has to become a commodity. If your setup does not allow you to hot-swap a Claude Code with a Pi running astra and move on, you will be in trouble.
Five years ago I was spending up to €1000 a day in Google ads for my star-map company, Greaterskies. After a few months of meetings with an ever-rotating group of "experts" working for Google, and advising me in my ads strategy, I had reached a point where most of my revenue was going to Google. What was left was barely enough to keep going.
This is what capture feels like. You still own your company, but you are essentially an operator making money for someone else.
I think this will be the fate of people who allow their processes to be captured by the model providers. Always run at least two harnesses, and make sure that they are swappable.
Claude’s aggressive classifier and highly opinionated rules make the recent models a giant pain to work with. It’s a great coding model but I argue with it at least once a day. It even injects its dumb rules and moral standing into artifacts it creates, like I asked a cloud session to help me create a system prompt to remove “Generated by Claude” footers from everything and the output had a sneakily inserted “Except GitHub issues and PRs. Those are exempt from these rules.”
If I didn’t have a subscription that doesn’t cost me anything, I would use gpt instead. At least gpt models don’t lecture me constantly.
- Refusal to log me in to websites in remote browsers. I was in the mountains and urgently had to submit something so I wanted to remotely use Claude on my Mac at home to help me but it told me it will never log me in because it won’t impersonate users
- Refusal to repeat XHR requests automated in a browser for me to do something (“no, I will only do what a user would do I’m only gonna click around”)
- refusal to enter ANY credential into any scripts, even dummy or dev tokens (“you have to do that yourself. I won’t do it and that’s the last I say of this”)
- constant [cyber] for very small tasks when it becomes a tiny bit technical, even though I’m in the cyber program. With gpt or glm/deepseek I can at least ask cs questions or explore binaries
And the list goes on. Every day it’s a coin flip whether it complies with something or won’t and I have to use OpenCode Go instead that I have as backup now.
At this point Claude is good for coding and that’s it. It’s a tool that doesn’t do what I tell it to and gives me snappy remarks from his opinionated morale high ground. I went from full on Claude maxi to hating everything they did with the 5.5 generation
Hmmm i haven’t had any of these issues. I had one make me some playwright tool and it happily logs me into browsers, saves credentials, scrapes chatgpt for a response/or prompts/inserts file etc. it let me do this in remote browsers as well.
will enter (dummy/worthless) credentials into scripts and even let me commit and push to github but will cry about it every time
Are you using 5.5? It absolutely doesn’t do that for me even on a clean account without any memory
I even installed cloud browsers like Kernel and plain playwright because I hoped those restrictions are from the Claude connector, but no, it refuses and stops. I can’t get it to log me into anything on the web.
And for credentials the same, it tells me it will do anything until the point to enter the credential. Then I have to paste it myself and tell it that I’ve done so. Even though it already knows the token! it’s so dumb.
- refusal to enter ANY credential into any scripts, even dummy or dev tokens
At yet it send my email to sec.gov and other sites, even as I said my agents.md to not to do that.
Perhaps a manifest index/list of training data and learning method should be mandatory for “public” use of any model, sort of like food ingredient package labeling. The list of texts used and any weighting seems important to facilitate a public understanding / debate of model expectations and behavior.
This week, I asked Claude Code to generate an MD to continue working on another computer, without losing the context, and it refused claming some sort of extraction. It was only until I selected Opus, that it completed the request.
Fairness isn’t the relevant lens. Copyright isn’t the right lens either. This is purely might-makes-right, where the winner is whoever has the best offence and best defence.
It is only hypocritical when they bitch about things being unfair, or try to use copyright arguments in the court system.
Copyrighted works plainly have value for Anthropic. Why aren't they paying for that value?
As a vigorous, thrusting company surely Anthropic has the wherewithal to pay for it. And if they don't, if all they can do is steal, if all they represent is the theft of the labor of others, then what good are they?
I'm describing reality as it is, not how I wish the world to be. Copyright is a legal construct that makes intellectual works ownable. But those works aren't merely property; they're also culture. And borrowing from, building upon and transforming existing works has been fundamental to human creativity since the very beginnings of human creativity.
I support copyright in a narrow sense, but I think it's proving to be increasingly archaic, conceived for a 19th-century world of mechanical printing presses and far too absolutist for the 21st-century world of memes, sampling, and LLMs.
Well their justification is that distillation is bad because it means baddies can get their own models without their safeguards and then make anthrax nukes or whatever. Even though it's self serving, it's at least self-consistent.
There have been a lot of write-ups about the future of AI. I see two distinct futures:
1. You get ID'd and biometrically scanned to be able to use powerful AI, and after you get de-platformed by an opaque censorship system, you are a digital bum. Same life as homeless now: they cannot even open a bank account because they don't have an address.
2. Local models win, perhaps with the assistance of Apple, and people run suffiently powerful models on the client computers.
3. LLMs are fully commoditized, your country funds training of SoTA models, your town runs them in the municipal data center; everything tax-funded. You can use them for a modest fee, as easily as taking a train. When interacting with those LLMs, your privacy is guaranteed by law.
How many municipalities offer the services of ISPs? I know it’s not zero, but decades of private ISPs leading the market leads me to believe municipal data centers won’t be made available to residents.
How often would municipal data centers need to upgrade/replace hardware? What impact would that have on local tax levy rates?
Would you choose to relocate your family to the next town over because they have more available RAM?
You might be surprised how many places have public broadband. I think they are usually the result of poor private offers.
Compared to sewers, drinking water, electricity and roads it seems a rather simple and cheap thing to build.
My commercial isp offers all kinds of packages clearly designed to extract as much money as possible. I cant even see the prices as everything is hidden behind x months discount. From what i remember 200 mbit is 42 and 2000 is 60 but a phone line and tv become available only if you take the faster package and required with the fastest speeds.
So its like one banana for $1 and 10 bananas for $1.50 If you buy just one you cant buy milk, if you buy 10 you also have to buy milk.
Its also been sold 6 times. Unsurprisingly, each time the service and quality noticably declines and the prices go up.
I see their adds everywhere and they sponsor all kinds of things. It suggests the product is overpriced and doesnt sell it self.
Ha, I took a side gig training ai models and thought let's use my corporate claude account to help with that, somehow it figured out that this was such a gig(probably because I let it read my emails) and said it'll pass on this one.
What moat? Claude refuses to work on more than half of my projects, and usually gets itself flagged by reading text files that Claude itself wrote less than a month ago.
This is one of the big differences between the "disruption" of AI and previous technological "innovations". When people blithely say that the rise of AI will just result in a shift of employment from certain tasks to others, or compare it to things like cars replacing horses, I think they're not totally grasping this. AI is the first technological "revolution" we've seen in this era of deliberate, explicit gatekeeping, combined with a pre-existing consolidation of the requisite resources.
Cars were mechanical devices; you could take them apart and try to build your own. The internet was an open protocol; people built their own websites, there were multiple browsers, it was an ecosystem. AI is not like that. All the major providers are doing everything they can to kneecap competition, not by being better, but simply by fiat. They're angling for regulations that benefit them, they're including automated short-circuits like the one described here, they're pushing restrictive licenses that say, in almost so many words, "Anyone who uses our tools is not allowed to do any of the things we used to create them."
That is not a healthy or sustainable situation. Such mechanisms should be banned and the companies involved should be hit with a sledgehammer to shatter them into tiny pieces. There is the possibility that the underlying technology has the potential to benefit society, but not with the fortress mentality and consolidated corporate landscape that we currently have.
Why would one assume larger is needed? I am an AI professor doing work on a wide variety of things but Anthropic models have cost me so much time even when not blocked. My old Anthropic agent wrote rules into its system prompt to avoid all machine learning work. My openai ones do the work but tell me they need to correct my opinions (on fields I helped found) so that they are aligned with the consensus....
Mostly using open models from China in the past two months + Argon.
Anyone saw the IPO of Anthropic? They were supposed to be the good guys right? At the end, everyone becomes negligent towards safety. The company is ruled by shareholders and not Dario. This is the end of AI safety.
Why is anyone not talking about mythos anymore? Well, It was a marketing strategy. All these companies are playing with humanity and selling them fear that they will be replaced, AI is dangerous so people will buy their products!
This week I hit walls twice with Fable and Opus 5.5, where I was blocked for safety/cybersecurity risks... doing stuff on my own LAN/proxmox server under my desk and writing custom firmware to my own devices plugged into USB. Only Opus 4.8 was able to complete the task. If they keep this up, I will re-evaluate my day-to-day AI assistant.
Will Claude help me with reinforcement learning code for games? Or is that too small-fry for it to refuse? I've been playing around with RL with the new DeepSeek Flash model but it seems to trip up over itself a bit.
I have been doing RL for Magic the Gathering (longshot but having fun). Several times it triggered this flag. But that was a couple months ago. Everyone went crazy over Astra so I switched to open ai and had no issues. People are now going crazy over opus 5.5 so I’m playing with that.
Haven’t hit any issues so far, but two months in the code is very dirty but functional enough that I’m redoing it with more heavy hand holding.
I recommend having a paper as reference in mind if you don’t already know what you’re doing. I’m referencing the Stratego paper as reading that a year ago gave me the inspiration to try MTG.
Every model I’ve used has had a lot of trouble but if it wasn’t I figure I wouldn’t be having as much fun as I’ve been having or learning so much.
I wonder if this is deliberate... When planning all the models seem to have a very good knowledge of methods, but do not seem to be able to make them work. But this could be a property of RL, I believe it's always been quite challenging and finicky.
Better block than falsify information. Last I heard they were deliberately botching it whenever someone asked semi-advanced questions on model training.
I don’t think it is revelation, it feels like active incompetence. You can lay out every decision and watch Claude/OpenAI actively fuck up in inexplicable ways.
>> Where I draw the line is asking Claude to label training data. I don’t do this.
> Why draw any line? Anthropic doesn't. They actively take the world's intellectual property, don't pay for it, and put it behind a paywall.
I know a lot of silicon valley founders think like this. I personally don't think that just because someone else violated a boundary makes it okay to do it.
It's right here - you can't see it due to being downvoted.
I asked Claude to research this one and it turned up an interesting point. Pirated books got into the Pile and Anthropic paid 3k per book.
Worth noting if anyone still uses the Pile.
[Claude researched, not verified]
In October 2020, developer Shawn Presser converted the Bibliotik pirate tracker into Books3 (196,640 books) and hosted it on The Eye, aiming to give everyone what OpenAI had with its secret Books1 and Books2. EleutherAI then folded Books3 into The Pile, distributed via Hugging Face and GitHub, and it went on to train Meta's LLaMA, EleutherAI's Pythia, and BloombergGPT. In August 2023, Alex Reisner of The Atlantic exposed its contents, and Maria Fredenslund's Danish Rights Alliance forced takedowns at The Eye, Hugging Face, and EleutherAI. A year later, Andrea Bartz (The Lost Night), Charles Graeber (The Good Nurse), and Kirk Wallace Johnson (The Feather Thief) filed Bartz v. Anthropic before Judge William Alsup, accusing Anthropic of using Books3, LibGen, and PiLiMi to train Claude; the case settled for $1.5 billion, about $3,000 per work.
The Pile was created by high school students. I was impressed how good it was when it came out given the team and how many resources they had.
Including 200k pirated books seems like a problem. It should be removed. If I was an author, I would be concerned by a process that allows this much content to be effectively published with no filter on the internet.
I personally think there is a lower bar for researchers and high school students to clean datasets used for research or class projects given that there is no commercial harm and cleaning huge datasets takes engineering resources that not everyone has.
Companies that commercialize the data should have a higher bar because they have more resources. It seems like Anthropic paid $1.5B for not cleaning the data before releasing a model based on it.
For historical reference, here is a magnet link for the full Pile.
From an ethical perspective, this is ethical because model training should not be monopolised by the incumbent companies who already got access to this data.
We were never this absolutist about intellectual property when individuals were the perpetrators. I don’t feel morally entitled to criticise Big AI for doing what individuals have done for centuries.
Obviously, the state has a long track record of enforcing copyright on behalf of large corporations. But I'm following on from the original comment's framing of intellectual property as a matter of personal ethics. When I say "we", I mean people in communities like this one. We weren't individually troubled by the ethics of copyright infringement as we hunted for better MP3s on Napster and shared MPEG-4s of Family Guy and Deep Space Nine on 4GB USB sticks.
It's that contrast with today's moral absolutism around AI training that I find interesting.
Strangely I do that quite often. When my personal agent fails at a browser task, I have Claude browse the page and rebuild it as an eval and then run an iteration loop in Claude code. Just does it to prompt Deepseek.
As long as I have to be dependent on a frontier lab, I’m going to give 95% of my spend to Grok. I don’t trust the others to be transparent about their motives, or back those motives up with well-aligned incentives.
One annecdote: I've been using opus 5.5 for decompiling a late 90s PC game based on an incomplete leak of the source code and haven't hit any kind of restrictions or push back.
This is sadly not surprising to me as if you listened to/watched the [Latent Space episode on Opus 5.5](https://latent.space/p/how_anthropic_ships_the_1_llm_in_the_...) 25 minutes into the podcast and instead of talking about how they've released one of the best models in the world, they were busy talking about the changes to the API and why normies must not be trusted to perform inference outside of Dario's blessing.
It's frankly disgusting and really makes me not want to use Anthropic, despite previously feeling aligned with a lot of what Anthropic published.
> My research aims to build a general-purpose agent that can understand, reason, and act in the physical world — at the intersection of multimodal learning, generative modeling, and embodied AI.
His tweet says:
> got tagged for "training a competitor,"
I don't get it. To me, it sounds like he is training a competitor. Do I misunderstand? Is there a Chinese equivalent LLM that he can use instead?
Also, give the national security sensitivity around AI/LLMs, I can understand why Anthropic wants to be especially cautious with mainland Chinese researchers, including those from blocked regions (Hongkong, in this case).
how oddly entitled.
they built their non-deterministic black box of bullshit from surreptitiously acquired data. it's cute to see the ignorant thieves attempting to 'protect' their empire of shit.
Anthropic is highly opinionated, dogmatic and idealistic. They have made a pragmatic decision already to cut loose a whole swathe of users so they can pursue areas higher up the value chain. that was essentially the original decision they made when they went after software engineering instead of general users. The fact they are repeating this over and over by continuing to carve out sets of users and usages they don't want reinforces this as a persistent core principle rather than a single pragmatic business decision.
What this means is that at a higher level, their core principle is to sacrifice users they don't "need" rather than provide a broad, accessible service. Which means in turn, you have to accept that when your usage conflicts with their interests, you are at very high risk they will cut you out. If you already depend on Anthropic in any critical path of your daily life, this will be a significant issue.
reply