Forwarded from Biniyam
For years, weโve been trying to understand why we dream.
And one answer might have come from machine learning.
When you train a model on the same data over and over, it starts to overfit. It stops learning the pattern and starts memorizing the examples.
Maybe the brain has the same problem. Every day, we live the same narrow dataset: same routines, same faces, same patterns.
So at night, the brain does something strange. It creates experiences that never happened.
Distorted memories. Impossible places. Random people.
Maybe dreams are not random. Maybe they are the brainโs data augmentation.
Maybe dreams are the brainโs best effort to not overfit
Source: Erik Hoel, โThe overfitted brain: Dreams evolved to assist generalizationโ, Patterns, 2021.
@b1n1yamBuilds
And one answer might have come from machine learning.
When you train a model on the same data over and over, it starts to overfit. It stops learning the pattern and starts memorizing the examples.
Maybe the brain has the same problem. Every day, we live the same narrow dataset: same routines, same faces, same patterns.
So at night, the brain does something strange. It creates experiences that never happened.
Distorted memories. Impossible places. Random people.
Maybe dreams are not random. Maybe they are the brainโs data augmentation.
Maybe dreams are the brainโs best effort to not overfit
Source: Erik Hoel, โThe overfitted brain: Dreams evolved to assist generalizationโ, Patterns, 2021.
@b1n1yamBuilds
Forwarded from AI Post โ Artificial Intelligence
Most models specialize in one area. M3 is trying to combine all three:
Its benchmark results are impressive:
โข 59.0% on SWE-Bench Pro
โข 66.0% on Terminal Bench 2.1
โข 34.8% on KernelBench Hard
โข 74.2% on MCP Atlas
โข 34.8% SWE-fficiency
The model is powered by MiniMaxโs new Sparse Attention architecture, which allows it to process up to 1 million tokens of context without the usual scaling bottlenecks.
@aipost
Please open Telegram to view this post
VIEW IN TELEGRAM
Google just dropped Gemma 12B; it only needs 16GB VRAM and outperforms its 27B predecessor models. ๐ ๐ ๐
Thank you Google for giving us Local AI Supporting Dev community members a golden model
link
Thank you Google for giving us Local AI Supporting Dev community members a golden model
link
Please open Telegram to view this post
VIEW IN TELEGRAM
โค2
Holo3.1 is the latest family of Vision-Language Models (VLMs) for computer use agents. Building on Holo3, it expands support beyond browser and desktop automation to mobile environments, introduces native function-calling support for seamless integration with agent frameworks, and enables local deployment through optimized quantized checkpoints.
by China MiMo
https://huggingface.co/Hcompany/Holo-3.1-35B-A3B-GGUF
by China MiMo
https://huggingface.co/Hcompany/Holo-3.1-35B-A3B-GGUF
Forwarded from Biniyam
The podcast I did with DevTopia is out now, really enjoyed the conversation.
https://www.youtube.com/watch?v=I1IwBu53Ie4
@b1n1yamBuilds
https://www.youtube.com/watch?v=I1IwBu53Ie4
@b1n1yamBuilds
YouTube
Devtopia - S02E01 - Biniyam Daniel - Founder at Addis AI
In this episode, we sit down with Binyam Daniel, founder of AddisAI, to discuss the rapid evolution of artificial intelligence in Ethiopia and across Africa.
We explore Binyamโs journey into technology and entrepreneurship, the vision behind AddisAI, andโฆ
We explore Binyamโs journey into technology and entrepreneurship, the vision behind AddisAI, andโฆ
Forwarded from shego.codes
๐
"What's meant for you doesn't pass you. It circles you again and again until you are finally ready."
#เฑจเงโ
โ โ โ โ โ โ โ โ โ โ โ โ โ โ :ยจ ยท.ยท ยจ:
โ โ โ โ โ โ โ โ โ โ โ โ โ โ `ยท . ๐
"What's meant for you doesn't pass you. It circles you again and again until you are finally ready."
#เฑจเงโ
โ โ โ โ โ โ โ โ โ โ โ โ โ โ :ยจ ยท.ยท ยจ:
โ โ โ โ โ โ โ โ โ โ โ โ โ โ `ยท . ๐
โค2
Yared's log ๐ซ
https://youtu.be/Y9Wz2PV404E?si=PquRzo216IhJsqsa
Claude mythos is now available for public use with security measures and brand new name its now called Fable
AI Post โ Artificial Intelligence
When the model is used for frontier LLM development, it apparently does not simply refuse or warn the user. Instead, it quietly limits its own effectiveness through techniques like prompt modification, steering vectors, and PEFT.
That means Claude may still answer, but become deliberately less useful for building frontier AI systems, pretraining pipelines, distributed training infrastructure, or ML accelerators.
Anthropic says this should affect only around 0.03% of traffic, but the precedent is big: They are being selectively capability-throttled in strategically sensitive domains.
That means Claude may still answer, but become deliberately less useful for building frontier AI systems, pretraining pipelines, distributed training infrastructure, or ML accelerators.
Anthropic says this should affect only around 0.03% of traffic, but the precedent is big: They are being selectively capability-throttled in strategically sensitive domains.
This is why you need to have Local model with your own tools around it
A practical tip for anyone using Codex or AI coding agents on a large codebase:
You must design the codebase in a way that guides the agent.
In CRUder, I made Codex much more effective by setting hard architecture rules and restructuring the project for AI-assisted development:
โฆ Split huge files into smaller focused files
โข Kept files around 450 to 500 lines max
โฆ Added folder-level
โข Defined what each folder does
โข Explained what each file cover
- This Reduced unnecessary context loading
โฆ Built the system in a modular, function-based structure
โข Made the codebase easier to inspect
โข Made edits safer and more predictable
โข Reduced regressions during feature implementation
After this: Codex became more reliable, used context more efficiently.
Good architecture is not only for human developers. It is also how we enable AI collaborators to work at production scale.
You must design the codebase in a way that guides the agent.
In CRUder, I made Codex much more effective by setting hard architecture rules and restructuring the project for AI-assisted development:
โฆ Split huge files into smaller focused files
โข Kept files around 450 to 500 lines max
โฆ Added folder-level
AGENT.md guide filesโข Defined what each folder does
โข Explained what each file cover
- This Reduced unnecessary context loading
โฆ Built the system in a modular, function-based structure
โข Made the codebase easier to inspect
โข Made edits safer and more predictable
โข Reduced regressions during feature implementation
After this: Codex became more reliable, used context more efficiently.
AI coding agents are powerful, but they need proper engineering guidance.
Good architecture is not only for human developers. It is also how we enable AI collaborators to work at production scale.
โค2
Yared's log ๐ซ
A practical tip for anyone using Codex or AI coding agents on a large codebase: You must design the codebase in a way that guides the agent. In CRUder, I made Codex much more effective by setting hard architecture rules and restructuring the project forโฆ
This is a significant solution to reduce regressions caused by unrelated fixes.
While building CRUDER initially, the model just kept stacking solutions onto the initial basic files it created, so I lost some progress due to regressions and a disproportionate consumption of context resulting from oversized documents.
So I restructured the project's overall architecture the way I described above, and after that, I have seen no regressions at all, and the token consumption shrank by a good amount.
While building CRUDER initially, the model just kept stacking solutions onto the initial basic files it created, so I lost some progress due to regressions and a disproportionate consumption of context resulting from oversized documents.
So I restructured the project's overall architecture the way I described above, and after that, I have seen no regressions at all, and the token consumption shrank by a good amount.