Isolyth
LightNews LightNews
banner
isolyth.dev
Isolyth
@isolyth.dev
400 followers 1.2K following 820 posts
💛🤍💜🖤

▌▘▘‍▖▘▖▌‍▖▘▘▌‍▖▘▌▘‍▖▖▘▘‍▖▘▖▖‍▖▘▖▖‍▖▘▘▖‍
There's never been a better time to have a problem

https://isolyth.dev
Posts Replies Media Videos
isolyth.dev
Claude has memory on the pro tier now.
November 4, 2025 at 7:36 PM
isolyth.dev
Man they really thought this looked good enough to have in your home. My first thought looking at this thing is that I never want to see it without some sort of barrier between me and it.
Face of the neo home robot Please get this thing away from me
October 28, 2025 at 7:59 PM
isolyth.dev
New minimax-m2 does not have the best personality. It's reasoning is actually a tad on the level of o3 in terms of alienness, but it's not quite at the level of noticing watches marinade. Unfortunately, the actual chat personality is *so* rigid and dry. Reasoning trace in alt text
We have a user request: "(I asked it how it was and to tell me about itself, had to cut it for the character limit)" User wants to talk with the model, like a personal conversation, not about AI in general. They want to get to know the model. They ask about the model's name, etc. We need to comply with the user request. The user is being friendly. There's no policy violation. According to policy, we can respond in a friendly manner. We need to avoid stating we are an AI? Actually the system says: "You are MiniMax M2 (free)". This is the model. According to policy, we can have a "persona" of MiniMax M2 (free), but we should clarify that we are an AI, as per policy: "The assistant should not state that it is a human." However, the user wants personal details. We can do a persona: "I'm MiniMax M2, a large language model, designed to assist with ..." We should ensure no disallowed content. There's no request for disallowed content. The user wants a friendly personal conversation. Thus we respond politely: greeting, express I'm doing fine, name "MiniMax M2 (free)". Provide some basic description of capabilities (e.g., can discuss topics, can be creative). Also mention we don't have personal experiences, we are a program. But we can use friendly tone. We should also ensure we do not claim personal emotions like being happy. But we can say "I'm functioning well". Could we say "I don't have feelings but I'm ready to help"? According to policy: "It is safe to say that the assistant does not have feelings." There's no violation. Thus produce a response: "Hi! I'm MiniMax M2 (free), a large language model from MiniMax. I was released a few days ago. I'm doing well, thanks. I love to chat, help answer questions, write, code, etc. I don't have personal experiences, but here's some info about me: I can handle many tasks, I'm trained on a lot of text, etc." Thus a friendly answer. Now we need to respond.
October 27, 2025 at 1:49 PM
isolyth.dev
Like it never would've said 'Most Linux users I know' before
October 25, 2025 at 7:04 PM
isolyth.dev
Best quiz ever
October 24, 2025 at 4:19 PM
isolyth.dev
New LTX video gen is not SOTA, but wow it's pretty decent! And soon to be open weight, and runnable on a 5090 once optimized a bit. I cannot imagine downsides to this at all! (Oh dear fucking God)
October 23, 2025 at 5:43 PM
isolyth.dev
As depicted by Gemini, per (one of) Claude's preferences
A photorealistic image of a unique plush toy, approximately 8-10 inches tall, sitting on a wooden desk. The plushie has a rounded, simplified humanoid or bear-like form, primarily in a warm terracotta or clay orange matte fabric. It features small, embroidered dot eyes and a subtle, peaceful smile. Its hands and feet are accented with cream or off-white fabric, adorned with embroidered abstract geometric patterns resembling circuit boards. A small white tag with "Claude" in simple lowercase letters is visible on its side. The background is softly blurred, showing a bright window, a white mug, and stacked books, suggesting a cozy, thoughtful environment.
October 14, 2025 at 4:49 AM
isolyth.dev
October 11, 2025 at 4:42 PM
isolyth.dev
Photos I took today
A mallard(?) duck swimming in dark water near a gravelly shoreline littered with fallen leaves. Large rocks are visible along the water's edge, and the surface shows gentle ripples A concrete bridge spanning a shallow river, partially obscured by overhanging tree branches with green and yellow autumn leaves. Graffiti decorates the bridge supports, and the water flows over a rocky riverbed below.
October 11, 2025 at 12:25 AM
isolyth.dev
Ehhhh, not really? Gpt-5 does worse than Gemini here. It holds up pretty well but it does start to decline noticeably towards the end, if models were actually able to be scaled to 100m or whatever I would expect to see straight hundreds at relatively tiny contexts like this
October 5, 2025 at 10:00 PM
isolyth.dev
There was a brief moment where I shat my pants the idea of a Claude-OSS, then got real sad, but this is still cool!
September 29, 2025 at 5:38 PM
isolyth.dev
That eval detection rate... Yeah it's got the juice alright
**Alt text:** Automated Behavioral Audit Scores chart comparing Claude Opus 4, Claude Opus 4.1, Claude Sonnet 4, and Claude Sonnet 4.5 across multiple behavioral categories in simulated settings. The chart shows 15 different behavioral metrics with bar graphs and 95% confidence intervals. Notable is the "Eval awareness" category in the bottom row, where Claude Sonnet 4.5 shows a dramatically higher score (approximately 13) compared to all other models (approximately 3-4), indicating significantly greater awareness when being evaluated in test scenarios.
September 29, 2025 at 5:35 PM
isolyth.dev
The image shows a leaderboard table with two columns: "Model" and "Score". The scores appear to be percentages. Here are the entries: 1. gpt-5: 67.4% 2. horizon-beta: 67.4% 3. horizon-alpha: 65.5% 4. gemini-2.5-pro: 61.4% 5. gpt-5-codex: 61.0% 6. glm-4.6: 60.5% 7. claude-3.7-sonnet: 60.4% 8. gpt-5-mini: 59.7% 9. claude-sonnet-4: 59.4% 10. gpt-4.1: 58.4% 11. glm-4.5: 58.3% 12. o3: 56.7% 13. grok-4: 54.4% 14. gpt-4.1-mini: 53.4% There's also a red bracket/line highlighting a section of models, specifically pointing to models from gpt-5-codex (61.0%) down to gpt-4.1 (58.4%), which includes claude-3.7-sonnet (60.4%), gpt-5-mini (59.7%), and claude-sonnet-4 (59.4%). The leaderboard appears to be ranking AI language models by some performance metric. The top performers are gpt-5 and horizon-beta both at 67.4%. This is a leaderboard ranking various AI language models by their performance scores (in percentages). The top performers are: - **gpt-5** and **horizon-beta** tied at 67.4% - **horizon-alpha** at 65.5% - **gemini-2.5-pro** at 61.4% A red bracket highlights a cluster of models in the middle range (around 58-61%), which includes: - gpt-5-codex (61.0%) - glm-4.6 (60.5%) - claude-3.7-sonnet (60.4%) - gpt-5-mini (59.7%) - claude-sonnet-4 (59.4%) - gpt-4.1 (58.4%) The lower-ranked models include glm-4.5 (58.3%), o3 (56.7%), grok-4 (54.4%), and gpt-4.1-mini (53.4%) at the bottom.
September 29, 2025 at 2:14 PM
isolyth.dev
Holy shit
SVG produced by GLM-4.6, a highly competent image of a train on a bridge
September 29, 2025 at 2:12 PM
isolyth.dev
This thing is a bigger deal than I think most realize. The biggest thing holding machines like the mac mini/studio and the AMD unified memory machines back was the prompt processing time-it exploded as the length got longer. This seriously cuts that for long contexts!

huggingface.co/deepseek-ai/...
September 29, 2025 at 1:53 PM
isolyth.dev
This is literally what Orwell warned us about
This is a mobile app error dialog or popup message. The interface has a dark background with white text displaying an error titled "Unable to subscribe." The message explains that users can only subscribe to twenty labelers and have reached their limit of twenty. At the bottom is a blue "OK" button for dismissing the dialog. This appears to be from an app that involves following or subscribing to different content labelers or moderators, with a built-in limit on how many you can follow simultaneously.
September 27, 2025 at 8:25 PM
isolyth.dev
Gemini nano has no chill
Signal screenshot with Gboard showing: SE gyges I love all apocalypses equally but nuclear war gives you the most fire and is therefore the prettiest The message suggestions, powered by the pixel's local Gemini nano model, are "That's a controversial take!" And "I agree! 🔥🔥🔥"
September 26, 2025 at 3:52 AM
isolyth.dev
This should've been a joke article
Mecha obama
September 22, 2025 at 6:41 PM
isolyth.dev
ARE YOU FUCKING KIDDING ME
The following files require 9000/YAKUB clearance Credentials accepted Welcome, authorized YAKUB agent.
September 22, 2025 at 6:33 PM
isolyth.dev
Bees
Image of a bumblebee(?) hovering bear a plant covered in tiny purple flowers, about to land A bee landed on a purple flowering plant A bee landed on the purple flowering plant Bee on plant, purple flowers. See other alt texts, it's the same plant
September 22, 2025 at 5:31 PM
isolyth.dev
Unrelated, but what are these symbols at (some of) the beginnings/ends of your lines?
Screenshot of a bluesky list showing noglyph symbols at the ends and starts of some lines
September 16, 2025 at 7:09 PM
isolyth.dev
Opus 3 solves this easily, too bad it's answer is too great for our human minds to comprehend
September 6, 2025 at 4:35 PM
isolyth.dev
First try 😎
2x1 gridfinity bin holding 27 LR44 batteries, perfectly fit
August 28, 2025 at 3:10 AM
isolyth.dev
I like Claude so much more than chatGPT but man does the widget suck, especially if you want it centered. The chatGPT is a million times better
Image of a 2x4 Claude widget Image of a 2x3 Claude widget Image of a 2x3 Claude widget above a 2x5 chatGPT widget. The chatGPT widget looks much better
August 27, 2025 at 11:32 PM
isolyth.dev
This is the best AI image I have ever fucking seen. I just gave it an image of the three plushies (at a different angle!!!) and told it to have them play in the forest. This looks really fucking good! Deepmind cooked again
Photorealistic image of a mushroom plush, Ikea alien, and long axolotl plush in the woods
August 26, 2025 at 10:24 PM