Loading activity…
@_chenglou
Worked on: @reactjs, @messenger, @reasonml, @rescriptlang Currently: @midjourney, Pretext, Freerange, Handwired
AI
A user shares an image link and jokes that the traces are “totally legible, nothing to worry about.”
AI
A user imagines storing data in memory instead of files, with excess data spilling to the cloud.
AI
Its creator says the net exactly matches JavaScript outputs, is fully differentiable and runs in a browser without a GPU.
AI
The creator says the JavaScript gallery uses backpropagation in real time and includes a slider for its loss parameter.
AI
Its creator says the process took 25 hours of Astra model/tool calls and over 27 rounds of human feedback.
AI
AcerFur used the line while quoting an earlier post, without explaining what “it” refers to.
AI
A user frames that possibility as “the death of closed source.”
AI
A post imagines 'Boris Claude' staying at Cursor and later making Grok Code at SpaceXAI.
AI
A user argues that treating AI like a person helps get rid of a task, while using it as a tool helps build expertise.
AI
A user asks about a large language model whose chain of thought—its step-by-step reasoning—must be executable math.
AI
A user fixing Pretext says Firefox uses an LSTM—a type of neural network—in ICU4X to determine word breaks for Thai and other languages.
AI
A post jokes about making a neural net “wordcell,” punctuated by “Who would’ve known.”
AI
A post quips that image generation belongs in the “transformer hole.”
AI
One post credits iOS's fluid feel to in-process animations and springs that preserve velocity. A related post recommends a Destiny engine talk as a model for extreme threading control, while noting most apps don't need it.
AI
One post jokes that robotics is “gonna become funny” once @typesafeai adds image input.
AI
The user says their fine-tuned model uses region tokens instead of treating every image patch as a token in every layer, while maintaining the same quality.
AI
The writer suggests the same person could approve or oppose an idea depending on whether its framing appeals to or repels particular groups.
AI
The post points to harness-design research as a way to improve performance, highlighting Graph-as-Policy (GaP), which it says was accepted to CoRL.
AI
“Some bugs are just harder than others,” the user wrote, saying on September 7 that a short update was coming soon.
AI
A paper linked by the user reports that repeating prompts improves performance for Gemini, GPT, Claude and Deepseek when they aren't using reasoning, without increasing generated tokens or latency.
