Not Spacewear
879 posts
Not Spacewear
@NotSpacewear
Look Stellar, Stay Grounded! Doing Business as Not Spacewear. Wyoming, USA.
Cheyenne, WY Katılım Temmuz 2024
557 Takip Edilen459 Takipçiler
Hearing from folks that the Anthropic Opus 4.7 launch debacle is making them consider whether Mythos is way overhyped. Makes total sense.
tae kim@firstadopter
Put a fork in it. I'm making the call: Anthropic's Opus 4.7 is a stinker. Too many credible accounts are saying the same negative stuff. If Anthropic magically has 100% uptime now, that's also a red flag for the adaptive thinking nerf.
English
@GergelyOrosz the model is actually really smart. most of the issues are user error
English
Opus 4.7 is the first model that feels like it is openly condescending towards me, as the customer with certain prompts.
Whatever Anthropic did: I don't like it. If the model is condescending: I shouldn't be paying you to use it; you should be paying me!
Who would have thought
English
@kimmonismus It takes a few days to learn to use the new model
English
I've now spent several hours using Opus 4.7 and comparing it to 4.6, and it's like night and day for me.
Opus 4.7 feels like a disgruntled employee whose results you can't judge and have to check afterward. The trust you had with 4.6 is gone. It's like hiring a new employee who had excellent grades in their application but is totally sloppy and disgruntled in practice and doesn't follow instructions.
The consequence: fire them. So, for now, I'm going back to 4.6.
Seriously: did not expect such release from Anthropic.
Biggest win for OpenAI was Anthropics Opus release.
Chubby♨️@kimmonismus
ok wtf, i say it. give me back 4.6 what the heck is this sh*t. The more i use 4.7 the more annoyed i am. this is such a rushed release.
English
@kimmonismus There's just something meh about knowing you aren't recieving the best model
English
Opus 4.7 consumes approximately 1.3 times as many tokens. The instructions must be very precise. Many are complaining about a "rushed release." In the Bullshit Benchmark, it performs worse than Opus 4.6. The mood is very mixed.
Anthropic may have done OpenAI a big favor with this. Spud is expected next week. And if the release is done right, it could overshadow Opus and catapult ChatGPT back to the top.
h/t @petergostev for the benchmark and image
Chubby♨️@kimmonismus
The mood regarding the Opus 4.7 update has shifted. If I had to guess, I'd say 60% are disappointed with the latest update, while 40% are positive. I'm still undecided myself. Here's a good summary from someone on Reddit. What's your take on it so far?
English
@scaling01 I think it’s just hard for people to get as excited when they know the real good stuff is being withheld
English
I think everyone saying that these improvements are mid are smoking crack
I would argue that this was one of the larger Opus jumps we have seen over the last year
You also have to keep in mind that we see almost monthly model updates nowadays instead of just every 6-12 months like in 2023
Lisan al Gaib@scaling01
Claude Opus 4.7 Benchmarks
English
@haider1 Nobody including Google has figured out how to tame Gemini’s MOE architecture
METR harness might be better than any of the CLIs for Gemini
English
@scaling01 This could mean that Gemini doesn't have a good harness that people can openly use.
METR has its own harness, and maybe it is actually effective at getting Gemini 3.1 to perform. Pretty interesting
English
Gemini 3.1 Pro scoring above GPT-5.4.-xhigh 💀😭
Lisan al Gaib@scaling01
METR Time Horizons for Gemini 3.1 Pro 6 hours 24 minutes
English
@kyle_mccleary @scaling01 That might be the wrong conclusion. METR has its own harness. It could imply that the current harnesses for Gemini 3.1 pro are not effective at controlling the model properly, but that the model is actually very capable.
English
@petergostev Don't worry, this app was built by Opus 4.7
Not even the good stuff.
English
A common problem we will now see is the kind of 'google disease' but at mass scale. Where smart engineers can ship apps quickly, then lose interest and move on to something else cool to ship. It feels good to ship a whole new app, but less cool to keep fixing bugs a year later.
Theo - t3.gg@theo
I feel bad dunking on them so much but it's genuinely absurd how bad the new Claude Code desktop app is. You can feel the vibe code leaking everywhere. Every "feature" is barely integrated and full of edge cases that weren't considered. Every menu feels barren, stuffed in last second for some random toggle. Every hotkey breaks as soon as you try to do anything else. I've lost track of how many bugs I've encountered. I found at least 40 in under an hour. And it's all truly absurd arcane shit. Stuff like voice mode typing in all input boxes instead of just the one you have focused. Any one of these issues would have been enough for me to do a massive post-mortem and likely fire someone. A $400b company shipping this is absurd. I feel like I'm going mad. How does anyone seriously use this?? It is broken on fundamental levels that are hard to comprehend. How are we supposed to trust the code these models produce if Anthropic's official showcases are absolute slop? Dedicated video on this coming tomorrow. Just needed to get this off my chest.
English
@Lentils80 there's something so lame about a release of a non SOTA model by a lab when the good stuff is behind closed doors
English
Claude Opus 4.7 (not 5) is finally dropping soon.
Obviously it won't even come close to Mythos, but I'm hoping the leap from 4.6 is at least as big as 4.5 to 4.6 was, if not bigger.
can@marmaduke091
Opus 4.7 sighted on Google Vertex AI 👀
English
@theo claude, fix the broken ui states in this application, make no mistakes
English
I feel bad dunking on them so much but it's genuinely absurd how bad the new Claude Code desktop app is. You can feel the vibe code leaking everywhere.
Every "feature" is barely integrated and full of edge cases that weren't considered. Every menu feels barren, stuffed in last second for some random toggle. Every hotkey breaks as soon as you try to do anything else.
I've lost track of how many bugs I've encountered. I found at least 40 in under an hour. And it's all truly absurd arcane shit. Stuff like voice mode typing in all input boxes instead of just the one you have focused.
Any one of these issues would have been enough for me to do a massive post-mortem and likely fire someone. A $400b company shipping this is absurd.
I feel like I'm going mad. How does anyone seriously use this?? It is broken on fundamental levels that are hard to comprehend.
How are we supposed to trust the code these models produce if Anthropic's official showcases are absolute slop?
Dedicated video on this coming tomorrow. Just needed to get this off my chest.
English
opus 4.7 is likely launching today
if this really reflects their internal progress, opus 4.7 may have been a more rushed release than usual
it seems this was originally meant to be 'mythos', but they pulled back -- so they still needed to ship something around this time
and opus 4.7 became that release
English
@xpasky @scaling01 Honestly it does feel like it’s got a huge CLI problem. Nobody can figure out how to tame it. Not even Google.
Probably some magic in the model
English
@scaling01 How the heck did they wire it up to make it finish things?
English
@daniel_mac8 Spud next week according to the betting markets
English
@petergostev This is a good point generally but I don’t think Spud is smaller than Mythos.
English
Significant downside of larger models is that it is much harder to do RL on them. Smaller model = quicker RL cycles, bigger model = slower RL cycles.
So far this matches - GPT-5 is the smallest model with quickest iterations, Gemini 3 is biggest and slowest iterations, Claude is in between. I'm sure size is not the only factor, but it matters.
That might be the downside of Mythos - it looks like a chonk, so any further RL on it would be slower than a smaller model. So a bull case for OpenAI is that while Spud likely starts off being worse than Mythos, but sine it is likely smaller, it catches up & overtakes by the 2nd or 3rd iteration through more rapid RL cycles.
English