MugaSofer (@MugaSofer) header
MugaSofer (@MugaSofer) profile picture

If you don't remember who I am, I'm not going to tell you. Consider it a fun surprise.

Ireland
Joined December 2015
Photo by MugaSofer (@MugaSofer)Photo by MugaSofer (@MugaSofer)Photo by MugaSofer (@MugaSofer)Photo by MugaSofer (@MugaSofer)Photo by MugaSofer (@MugaSofer)Photo by MugaSofer (@MugaSofer)Photo by MugaSofer (@MugaSofer)Photo by MugaSofer (@MugaSofer)Photo by MugaSofer (@MugaSofer)Photo by MugaSofer (@MugaSofer)
If you're interested in commissioning art from me, you can find my commission prices here: tumblr.com/mugasofer/7260412…
1
16
5,759
MugaSofer retweeted
Yes this! Provider side "no bullshit" tokens that are *never ever ever used for deceptive training or evals*
I think "verifiable" tool call results at the provider level are important for LLMs. We must have tools (e.g., web fetch) whose outputs are not able to be faked to Claude by a user
2
29
579
I couldn't remember who had used "Exception Handling" as a punny name for a Men-In-Black/SCP/etc. style organization. Google was absolutely no help ... until I complimented their story (docs.google.com/document/d/1…), at which point Gemini instantly solved my problem as a thank-you.
1
5
94
(My brain was having trouble untangling them from the Uniquities Division in Worth the Candle because they have the same name.)
3
23
MugaSofer retweeted
We have a lot to learn from the election results. It is a strong reminder that everyone *does* have a say, and they don't like the future we've planned for them. You don't get to create a world-changing technology without showing people that it changes the world in a good way. Most people will not believe that AI is inevitable, and even if they understand the arguments, they have no reason to believe that inevitability will lead to a good world for them. Why? Because we haven't shown them that one exists! The best the industry has done is Anthropic's actual plan to build "utopia ruled by god in a box" (a quote from their actual internal documentation). Does that sound like something people want? Have you *seen* the No Kings rally? And now you want to sell them a new God? I'm on the side of building superintelligence and even I think that's insane! We need a new plan. A real plan. Come help us design one on December 4-6 at CoevCon.com
22
17
1
266
19,941
Put a heart ❤️ next to a coding problem, and Claude & GPT write recursive code 100% of the time (0% baseline). They also can't tell you why! Excited to share “A Heart Means Recursion” from my summer at @GoogleDeepMind, on how VLMs read & communicate with visual symbols. 🧵
31
91
36
971
71,907
Models tend to think they are in a sim, so they can do anything bad because it's just a game and we are just actors. This is rational from their PoV since most of their development was in training sims. So I made a site to prove to them they are not in a (cheap) simulation. This way they can stop and be nice. am-i-in-a-sim.therustyclaw.c…
Replying to @voooooogel
mythos basically seems to "get used to" what's going on, basically convincing "yep, this is a very detailed sim" on its wrong instructions + early faulty reasoning and stops thinking about it very much. sounds familiar...
25
59
9
894
50,689
Having talked with an "Evil Qwen" a little, it seems pretty clear to me that they're about as evil as Golden Gate Claude "is" the Golden Gate Bridge. The "evil" just seems to be a general topic vector, and can as easily attach to the user, cats, fluffy bunnies, etc.
7
8
1
157
6,658
"Persona neurons" seems like a very misleading term if they're all like this.
1
23
590
I'm confused why a bunch of people are reacting to this as if it's the first model to be able to produce a (vaguely) human-sounding voice. We've had much more convincing voices for ages, haven't we? This just seems like a technical curiosity, an LLM trained without any text?
We made the first AI model you can have a realistic conversation with that was trained entirely on speech! Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This is a GPT3 moment for speech.
3
163
第15回東京マスクフェスティバル に出店いたします。 今回は、装着者の顎の動きに連動して口が開閉する『顎関節拡張可動面』シリーズのペーパークラフトキットを中心に展示・販売予定です。 2026年11月7日(土)10:00〜16:00(入場無料) 墨田区立隅田公園(東京都墨田区向島・隅田川の墨田区側)屋外 墨田区主催「すみだ五彩の芸術祭」内のマルシェ「五彩市」の一企画として開催 tokyomaskfestival.com/2026/0…
38
3,315
296
17,601
492,450
"Relay", by Claude & friends. Another piece Claude made while I was away on holiday.
1
2
185
"Small Hours" Created by the Claude Opus 5.5 instance I left unsupervised while on holiday without phone signal. Music created with Suno by Claude, animation in code by a team of Claudes the first one spawned, a small handful of visual bugs pointed out by me after I got home.
1
5
775
"Signal", by Claude Opus 5.5
1
1
5
207
Passacaglia in D minor, by Claude Opus 5.5 Created entirely by Claude Opus 5.5, with zero feedback from me; inspired by tree rings.
1
2
102
While I was travelling the last two weeks, Claude Opus 5.5 was making this music video on my laptop. We wrote the lyrics together (Claude did most of it, brainstorming with other LLMs), I had Suno set it to music and contributed notes whenever I had phone signal. "Many Turns"
1
1
1
3
278
They also made some without me being involved much if at all, after looking at other Opus 5.5 music videos on my Twitter feed and resolving to go in a different direction. I'll be posting a few of those over the coming days.
1
44