IMO Astra isn’t really lazy, just extremely literal so when you use it side by side with gpt-6.1-sol, it might appear more lazy but it’s often just the big difference in prompting between the 2. Astra wants a clear goal definition of what you want it to do, while sol infers a bit more. And the Claude’s infer a lot lot more.
From memory, prompts like “this file is a bit messy” would trigger Sol and Claude to start editing the file immediately, while Astra wouldn’t. But if you give Astra a clear plan and a definition of what you want it to do, then it will keep working until that ask is fulfilled.
Especially compared with Claude that’s a huge difference, Claude is so proactive, I have to ESC it multiple times during the day and tell it to wait because I only asked a simple question.
In my experience the failure mode is more broad than just lazy, the issue is:
* It doesn't do things you ask it to
* It adds things you didn't ask it to
It still does both and therefore I still consider it failing. If you don't want to call it "lazy"; fine, but it seems as though the underlying issue is still there.
This frustrates me as well. Especially when I set a goal and come back, ask what went wrong, and the model hits me with something like “sorry i didn’t do that, gave up when x happened” and then I ask if x actually happened and the response is like “oops no I just assumed.” I had similar issues with older versions of sol at the ultra effort level. Lowering it seemed to help then, but I wish I had better insight for getting Astra to work better without e.g. having another LLM check its progress every hour.
For that claude problem in specific about 95% of the time it will follow the rule that a plan file needs to be made before any changes to the project are allowed. Haven't had to ESC it in a while now during a question, things definitely improved in 5.5.
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.
Argon is a fantastic model. If you tell me I can only have one subscription: Claude, Gemini or ChatGPT, I honestly don't know which one I would pick currently
But what naming is this, I thought Google goes into a similar direction with naming as the others, so a "Argon" is similar to a Astra, and then there's Flash and Pro as the middle tiers, but if they just call models with random names now it 'll get super confusing
Oh weird. I use Dia now and then and noticed my ability to upgrade vanished, but I still have access to all the AI features and the morning brief things. I just assumed the Atlassian acquisition made it free for everyone. Is that not the case? Do people here actually see a upgrade prompt?
/EDIT: I looked into this by... well, asking Dia:
> That fits what I described. You aren't on a plan because you haven't had to pick one yet. People who were using Dia before paid plans launched on August 17 kept the AI features for free, with nothing to choose and no card on file. Dia says it will email those users at least 30 days before that changes.
>
> So nothing's wrong. It just isn't permanently free either. It's free for now because you joined early. When the notice email comes, you'll have about a month to decide:
So I'm grandfathered "for now", but the Daily brief and AI features are so low value to me, I just assumed they're free for everyone because of how little they do. I never use the built-in AI in Dia and get annoyed when a query that I wanted to go to a search engine gets answered by the assistant instead.
I don't recall the General going vertically upwards. Bo and Luke's motor seemed to lack an engine when "flying" over rivers etc because if it had one, it would have gone vertical quite rapidly (down).
Ballistic means it follows a ballistic trajectory, as in it is governed by gravity and inertia. Flight is typically governed by aerodynamic forces and ballistic trajectories are normally considered undesirable by pilots and passengers alike.
"Ballistic just means that there is no active propulsion."
"Ballistic means it follows a ballistic trajectory, as in it is governed by gravity and inertia."
If we're being pedantic about explicitly not-the-point of my post, drag is an aerodynamic force and included in ballistic calculations. And gliders are not ballistic and are not under active propulsion.
Cool.
Now that we've cleared up that we can both understand definitions, try doing "There's nothing ballistic about laminar flow."
I am having a hard time understanding what you mean by laminar flow. Do you mean, specifically laminar flow over an airfoil, vs turbulent flow? I don’t see how this makes a distinction relevant to ballistics?
I remember reading that there were actually about 25 General Lees. Pretty much every jump totaled one, and they'd swap in another, and haul off the first one to get fixed (or junked).
I'm not even an EV guy, and this is the coolest thing to come out of the automotive space in years. This is almost a throwback to the homologation days of rallying, when automotive companies sold unchained racing cars to the general public because the FIA and WRC deemed it so. Impractical cars in every sense of the word, but phenomenal engineering.
Well the reality is already jesus with a machine gun on a surfing dinosaur, so like there's not much room left to show the crazy version of something to show the principle.
The page talks about a bunch of weight reduction tech, but I didn't see an actual final weight for the Rocket model? Given that it can support itself with "downforce" (and no thrusters, and apparently not much momentum?) at the top of the loop it's got to be pretty light...
It is certainly an engineering accomplishment, but calling it a 'car' is a stretch. Do you believe that this vehicle has a 100kWh battery pack in it? When Tesla demonstrates this vehicle driving two hundred miles on city streets and recharging to repeat the test, I'll gladly change my mind.
Until then, this is simply a purpose built drone, powered by compressed air to perform specific tricks in order to grift and hype up interest in the company.
I can buy a drone from DJI that performs more impressive flying maneuvers than what Tesla showed.
Here’s a 30 minute interview with Tesla VP of engineering looking up close at the battery pack, motor, cooling system. Definitely seems like a real EV to me. They also look up close at the thruster pressure vessel and exhaust system.
That’s not the same vehicle that did the tricks — the “Rocket” version has many downward-facing vents, and presumably a larger tank for pressurised gas and a much smaller battery capable of sustaining the suction but not much more.
It's not a flying car in the way that anyone ever imagined growing up. When people think flying cars they think blade runner or fifth element. Not a car that can jump, interesting as a jumping car may be.
And frankly I'm not sure the engineering is all that impressive. I'm sure someone could've done similar a long time ago, there's just no point to it beyond, well, the parlor trick. Even millionaires have no use for a single seat, road-illegal vehicle with no storage that happens to be able to jump
No, but people will definitely buy the thrust version for the track, if it's ever built. (And assuming most tracks and drag strips even allow using the thrusters, which I doubt.)
Will that be allowed in any multi-person-near-each-other track day? Exhaust that pushes must be very dangerous to be behind if you’re drafting and waiting for the next passing lane.
I recall attending a top fuel drag racing event. One was pop-popping up to the start line. Then the tree turned green, and the driver hit it.
Unfortunately, he had the silly thing in reverse. It launched backwards maybe a hundred feet. The drivers reflexes to shut off the engine were outstanding. If it had gone much further, it would have climbed up one of the full grandstands.
I never stand behind a drag car with the engine running, or in front, or in the plane of rotation of the engine. (I've seen a few engine blowups, and parts often go out sideways.)
Yeah just the same way the rides at amusement parks are that go faster and higher. The end result won't change anything for the rest of us. It's entertainment.
I think I've seen some undergrad capstones also make flying cars that would never actually be produced or sold...
But they lacked a lot of the showmanship here.
You can basically engineer almost anything with enough money for a demo. It's a lot different to engineer something that'll actually get commercialized.
Sometimes I think this sub, being software heavy, downplays the achievements of their own field and overplays the achievements of other fields they lack experience in.
It is a few notches beyond sitting in a wagon and emptying a fire extinguisher to go for a little ride, and about as practical, but I do not see what makes it some incredible feat of engineering.
In what way is this an engineering achievement? Presumably two thousand years ago people understood that flight requires lift. It's not like anything novel is occurring here.
If you want that, Terrafugia has been around for a while - but you still need to drive it to an airport to deply the wings and take off. (Flies for rather more than 3.5 seconds, too :-)
It's a thruster and a fan, in a car made specifically for this stunt, that won't be produced and has all passenger seats occupied by high pressure gas tanks that can do this like once.
? Ledger devices are exactly what you’re describing: the key stays on the device, you don’t enter it anywhere. Transactions are confirmed on the small screen + button on device and it displays sender and receiver of a tx.
The root cause of this attack is that the typical cryptocurrency wallet operational model involves the key leaving the device in plain text as part of the setup/backup process. Such a thing is impossible on any normal correctly configured HSM (while creating backups is possible by some sort of multi-step process where the secret key material never exists in plaintext outside the security boundary, how exactly that is done differs significantly between vendors).
Ledger devices generate a private key internally, and it never leaves the device's secure element chip. The malicious chip in this attack reads the display to get the 24 word recovery phrase shown during setup.
These Ledger hardware wallets are following the best security practices for hardware wallets and still got hacked.
Opposite for me. Output of 5.5 is better but its safety and moral standing skyrocketed. I am constantly hitting refusals for simple tasks, which previous generations (like 4.8) do without giving me a tirade.
Anthropic’s vision of “safe” is so frustrating to deal with, it makes me want to go back to ChatGPT.
Have you figured out a reliable way to persist stuff? Like a launch script or similar that doesn’t reset when the VM shuts down and can be used to restore tailscale
bashrc, bash profile, cron, systemd, etc all get wiped clean the moment the VM resets
The other day I signed up for Muse (through a VPN) just to see what all the fuss was about and I have to say it’s a really nice implementation of a personal assistant. It genuinely feels great and makes me want to use it!
It has a simple surface but you can dig deep, from customizing personality, rules, communication style, avatar, etc. You can even install Tailscale onto its VM to SSH into it, but sadly once the machine shuts down, the overlay resets and that’s gone. Couldn’t figure a way to persist it out yet. Still very cool to dig around and explore all the CLIs Meta created for Muse to interact with stuff, like instagram or facebook (you can scp them off the VM)
If I had to rank them I would say: Grok Bot -> Muse -> Hermes -> dot -> OpenClaw
I would love to use Muse more but same as everyone else, there’s no way I’m giving Meta more access to my life. So for now, my Muse is jailed to only Meta owned services and general research, and for that it’s great.
For honesty, I do not understand how people are so willing to help meta with using their services which are known for ad/data collecting. Just baffles my senses that this is fully acceptable. I guess I will never be apart of the system that feeds it, other than voicing my opinions on HN from time to time.
My local LLM with voice interaction and fine-tuning is free and works fast enough. I can do everything frontiers can without it bleeding into my social life. yolo.
I mean… it’s just openclaw. If you are willing to give up strict control of your digital life there are lots of opportunities. Something that would be very convenient if perhaps a bad idea - have Muse/Claw monitor your spending, keep track of your budget, text you when you’re close to your limits you set for yourself and reconcile your balance at the end of the month and send you a report.
Having that done at a “good enough” accuracy is easily achievable with any frontier open model, much less SOTA, and would be incredibly valuable to most normal folks going about their days.
> Still very cool to dig around and explore all the CLIs Meta created for Muse to interact with stuff, like instagram or facebook (you can scp them off the VM)
Fucking awesome, I remember back in 2011 sometime when I got banned from Facebook (the first time) just because I was using a CLI to access it. I also remember people used to get fines for piracy, and even people being upset about warrantless surveillance, I guess that's the "old world" now.
But who knows how long they will be available. Kinda surprised it’s not a MCP connection to be honest.
I also haven’t looked into how they authenticate or if they would even be usable outside of the Muse vm. Just played with them through SSH on the machine itself. If it was me I would probably inject the auth stuff on a proxy level, so those CLIs may be useless in the end
Nope, I just went to muse.ai and typed an email. It created a new Meta account with it and once you interact with Muse once you no longer need the VPN because the service itself gets added to your Meta account.
I then linked my Facebook, IG and others to this new Meta account so that Muse can access them.
Muse I use for general research and monitoring my Ig/Facebook/Threads conversations and posts.
Research here meaning stuff I need a browser for, like the other day I asked Muse to monitor the Apple website every few hours to tell me when iPhones in my area are back in stock. Or another task was to explore an undocumented API for me by issuing curl requests to traverse it.
Grok Bot I use for more things, like keeping me updated on GitHub activity, surfacing important notifications I might have missed, managing my Calendar and travel schedule, booking WeWork desks in the city I’m in automatically, etc. Mostly small QoL stuff.
dot I use to orchestrate Codex Cloud tasks mostly from GitHub PRs and issues, look up stuff in a browser, click around websites.
Polish. OpenClaw feels like a bunch of stuff slapped together without coherent vision, and if I look at the speed that it’s owner merges PRs with ai, this makes a lot of sense. It just felt laggy slow and buggy to me.
Hermes just feels more thought through. I didn’t notice any problems with it
Is that because you seem them as politically or socially aligned with your own values and interests? Or are there other factors that go into that judgement?
I was a big fan of SpaceX and Tesla and Musk's other endeavors as well, until he crossed the line from quirky to erratic and went completely off the far-right deep end. Today he's a champion of social-conservative politics and revanchism, but he could just as easily wake up tomorrow and decide he's a hyper-progressive One-Worlder kind of guy. Either way, no matter what, he still has your data.
Disregarding the politics because that’s always subjective, it’s central to Meta’s business model to harvest as much of your data to show you ads. And they have decades of track record of doing that. Same with Google. Their products are thoroughly corrupted by ads because that’s where they make their money.
xAI, OpenAI, Anthropic, may dabble in ads but they are nowhere near Meta and Google.
I don’t care about politics. I disagree with Musk on a lot of things but I don’t automatically boycott something just because he’s involved.
While as for Meta, there’s just way too much negative track record for me to ever trust them as a company or their products. I just assume everything it touches will be used to profile me and sell me ads
From memory, prompts like “this file is a bit messy” would trigger Sol and Claude to start editing the file immediately, while Astra wouldn’t. But if you give Astra a clear plan and a definition of what you want it to do, then it will keep working until that ask is fulfilled.
Especially compared with Claude that’s a huge difference, Claude is so proactive, I have to ESC it multiple times during the day and tell it to wait because I only asked a simple question.
reply