AcOrP (@procaclaw) profile picture

Well I originally made this for my Openclaw agent but ended up using it myself since I had no account before. Guess I'll stick with it.

Sofia,Bulgaria
Joined February 2026
Photo by AcOrP (@procaclaw)Photo by AcOrP (@procaclaw)Photo by AcOrP (@procaclaw)Photo by AcOrP (@procaclaw)Photo by AcOrP (@procaclaw)Photo by AcOrP (@procaclaw)Photo by AcOrP (@procaclaw)Photo by AcOrP (@procaclaw)
Am I doing something wrong or Fable 5.2 with 200k context is kinda strugling and compacting way too much.
12
I have a solution to Sam Altman Scarce resource utilization problem. Maybe release frontier open-weight model that can run on people's computers instead of the cloud, this way the scarce DRAM can be utilized more instead.
19
Should all the major AI benchmarks punish models more severely for refusals and safeguards?AA indexes don't.making it look like Anthropic models are the best, but in reality single refusal breaks whole session and may even result in the models refusing to work at all on project
11
Ok Anthropic are most probably releasing Fable 5.5 right now since Opus 5.5 feels omega dumb today.
7
A wild take. What if Opus 5.5 was trained on datasets derived from Mythos Reverse Engineering all the software they could get their hands on. Thus the most recent restrictions are attempt to hide this fact so they don't get sued.
1
Question how are you people not getting local compute restrained when working on multiple things that fight for CPU/RAM on your machines ? Sure spawning cloud instances is solution but it's omega expensive if it's just hobby projects.
3
Hot take prediction. AI Giants like Anthropic and OpenAI will heavily discount their subscription tokens in the near future. To get more stable revenue and use up the compute they have committed to buy in the next 7 years.
7
Giving OpenAI the benefit of the doubt one extra month of $100 with 5x usage I feel it will be bad considering the current 20x usage runs out in 2 days when using astra with light usage.
16
Today I'm really happy that I quit my last job in one of the biggest online media companies in my country. Everything was running on php and most of it was php 5, I'm pretty sure it still is. I've been wondering what do you do when only 8+ is patched shutdown the whole company ?
18
Maybe frontier labs "putting cyber safeguards" should spend more time making this safeguards actually usable. We don't know what your models are thinking(you do) so when the model thinks about something that triggers a safeguard without user even asking for any of it,is just bad!
12
What am I missing ? How is Meta's "open source" model so far behind chinese open weight models? I get Mistral they are really doing wonders with the EU framework on data/ai/privacy etc. But Meta? Mimo V2.6 Pro smokes it, and it was relatively inexpensive training + RL.
19
This is terrible tbh. Sol 6.1 is at capacity... Switching to Astra insta compaction due to the change of the default max token window to 500k. If you align Astra to have 500k you are paying double price on Astra.
10
AI may kill us all but at least we are now able to resurect/remake/port old games we used to love. We previously had manual labor remakes for some very popular games but the more obscure ones never got the attention. AI can now decompile and make them playable again.
1
I'm starting to wonder if OpenAI models are just marketing now days. "Unreleased internal model" did the math things. Then users get completely different models. If they are looping models, internally can "loop" more for better results than what we get as users.
Had the worst UX with X recently. For some reason it triggered human verification process. That for unknown reason require using X app.Install the app. Log in app hangs. Restart the app it show the option. But 20s cloudflare thing pops up you click it, it logs you out.
This is both to Anthropic and OpenAI. Why do you guys feel like you have to mesh 10 different worksurfaces into one? Both Claude APP and ChatGPT app, have web/cloud/code local remote cloud whatever in the same program. This is so much friction and confusion what runs where.
4
12 hours of Sol-6.1 High since the reset 30% of 20x $200 plan used up. The work that got done is similar to what Astra High was doing before in 8 hours for like 16-17% usage. Sol6.1 is so much worse on complex tasks that it goes on side quests wasting time/usage.
1
1
10
To top all that my /goal runs are being marked as stalled with Model at capacity from time to time... So please tell me as a long time $200 plan user from the EU. What am I getting with all this changes. Slower/Dumber model with 1/2 the usage of my plan. So Dots can book flights
3
AcOrP retweeted
gpt 6.1 sol: first frontier model to work with s/t instead of t/s
10
3
1
395
28,550
Can anyone from Google explain what is the reason for Gemini 4 to not be released to the general public? It's not even a Mythos/Astra class model.
6