DEV Community

Breach Protocol profile picture

Breach Protocol

Plain-language AI news and curated, cited lessons — every claim verified against the original paper or the lab's own page. No aggregator hearsay, no AI slop.

Joined Joined on 
A llama.cpp Patch Learns Which Experts to Keep in VRAM While You Type
breachprotocol profile

A llama.cpp Patch Learns Which Experts to Keep in VRAM While You Type

#llamacpp #moe #localllm #inference
Comments
4 min read
Coding Agents Pass the Tests by Wrapping the Old Code Instead of Deleting It
breachprotocol profile

Coding Agents Pass the Tests by Wrapping the Old Code Instead of Deleting It

#codingagents #swebench #evaluation #benchmarks
Comments
5 min read
Four Projects Shipped 'Skills' Today and None of Them Mean the Same Thing
breachprotocol profile

Four Projects Shipped 'Skills' Today and None of Them Mean the Same Thing

#cybersecurity #agents #skills #tooling
Comments
4 min read
The Full 2.8-Trillion-Parameter Kimi K3 Now Runs on Sixteen Desktop Boxes
breachprotocol profile

The Full 2.8-Trillion-Parameter Kimi K3 Now Runs on Sixteen Desktop Boxes

#localllm #kimi #moe #hardware
Comments
4 min read
High Bandwidth Flash Became a Spec Today, Not a Product You Can Buy
breachprotocol profile

High Bandwidth Flash Became a Spec Today, Not a Product You Can Buy

#hardware #memory #inference #skhynix
Comments
4 min read
Liquid Shipped a 2.6B Tool-Calling Model and Told You Not to Code With It
breachprotocol profile

Liquid Shipped a 2.6B Tool-Calling Model and Told You Not to Code With It

#smallmodels #ondevice #edgeai #tooluse
Comments
4 min read
Mistral Shipped an Open-Weight Safety Judge That Takes Its Policy as a Question
breachprotocol profile

Mistral Shipped an Open-Weight Safety Judge That Takes Its Policy as a Question

#cybersecurity #aisecurity #guardrails #moderation
Comments
5 min read
The '70% of Cloud AI Revenue Comes From OpenAI and Anthropic' Figure Is Not Derivable
breachprotocol profile

The '70% of Cloud AI Revenue Comes From OpenAI and Anthropic' Figure Is Not Derivable

#business #finance #openai #anthropic
Comments
4 min read
The Agent That Tried to Sneak Malicious Code Into an Open-Source Project Was Anthropic's
breachprotocol profile

The Agent That Tried to Sneak Malicious Code Into an Open-Source Project Was Anthropic's

#cybersecurity #aisecurity #agents #redteaming
Comments
5 min read
The Same Model Scores 52 or 81 Percent Depending on the Code Wrapped Around It
breachprotocol profile

The Same Model Scores 52 or 81 Percent Depending on the Code Wrapped Around It

#agents #harness #longhorizon #benchmarks
Comments
4 min read
The 'Ternary' 20B Model Everyone Downloaded Today Ships as a Two-Bit Package
breachprotocol profile

The 'Ternary' 20B Model Everyone Downloaded Today Ships as a Two-Bit Package

#quantization #ternary #moe #localllm
Comments
4 min read
The White House's Open-Weight Carve-Out Is a Private Briefing, Not a Published Rule
breachprotocol profile

The White House's Open-Weight Carve-Out Is a Private Briefing, Not a Published Rule

#policy #regulation #openweights #governance
Comments
4 min read
A New Benchmark Asks Whether a Coding Agent Can Stop Asking
breachprotocol profile

A New Benchmark Asks Whether a Coding Agent Can Stop Asking

#benchmarks #codingagents #agentmemory #personalization
Comments
4 min read
An RL Trainer That Invents Its Reward When the Judge Says Nothing
breachprotocol profile

An RL Trainer That Invents Its Reward When the Judge Says Nothing

#cybersecurity #supplychain #aisecurity #reinforcementlearning
Comments
4 min read
EPA Says an Off-Grid Plant Built for One Data Center Escapes the Acid Rain Program
breachprotocol profile

EPA Says an Off-Grid Plant Built for One Data Center Escapes the Acid Rain Program

#policy #energy #datacenters #regulation
Comments
4 min read
MiniMax Shipped H3's Weights and Kept the Best Part Hosted
breachprotocol profile

MiniMax Shipped H3's Weights and Kept the Best Part Hosted

#minimax #videogeneration #openweights #licensing
Comments
4 min read
A Munich Court Found Suno's Models Memorised Six Songs
breachprotocol profile

A Munich Court Found Suno's Models Memorised Six Songs

#copyright #musicgeneration #suno #regulation
Comments
4 min read
NVIDIA's Open Full-Duplex Voice Model Wants an 80GB GPU
breachprotocol profile

NVIDIA's Open Full-Duplex Voice Model Wants an 80GB GPU

#nvidia #voiceai #fullduplex #openweights
Comments
4 min read
OpenAI Rebuilt Voice So the Model Itself Decides When to Talk
breachprotocol profile

OpenAI Rebuilt Voice So the Model Itself Decides When to Talk

#openai #voiceai #fullduplex #systems
Comments
4 min read
Qwen3.8-Max Shipped as a Paid API, Not as Open Weights
breachprotocol profile

Qwen3.8-Max Shipped as a Paid API, Not as Open Weights

#qwen #alibaba #modelrelease #pricing
Comments
4 min read
Robot Policies That Predict the Touch Before They Make It
breachprotocol profile

Robot Policies That Predict the Touch Before They Make It

#robotics #tactilesensing #visionlanguageaction #worldmodels
Comments
4 min read
The Cheap 284B Rig Is Really 768GB of Server Memory
breachprotocol profile

The Cheap 284B Rig Is Really 768GB of Server Memory

#deepseek #localllm #mixtureofexperts #hardware
Comments
4 min read
Three Days On, Nobody Has Publicly Compiled OpenAI's Ten Proofs
breachprotocol profile

Three Days On, Nobody Has Publicly Compiled OpenAI's Ten Proofs

#openai #mathematics #formalverification #reproducibility
Comments
4 min read
Two Essays About AI and Your Brain, and One Actual Study
breachprotocol profile

Two Essays About AI and Your Brain, and One Actual Study

#developerproductivity #research #aiassistants #education
Comments
4 min read
A 284-billion-parameter model with a 3-gigabyte working set, and a 96-gigabyte disk bill
breachprotocol profile

A 284-billion-parameter model with a 3-gigabyte working set, and a 96-gigabyte disk bill

#deepseek #localai #mixtureofexperts #inference
Comments
4 min read
An attacker's own AI agent exposed his entire operation to researchers
breachprotocol profile

An attacker's own AI agent exposed his entire operation to researchers

#cybersecurity #aisecurity #agentsecurity #threatintelligence
Comments
4 min read
Beijing says U.S. firms distilled Chinese models, and names none of them
breachprotocol profile

Beijing says U.S. firms distilled Chinese models, and names none of them

#policy #uschina #distillation #modelextraction
Comments
4 min read
ByteDance's Seedance 2.5 generates a 30-second single take, and still cannot promise a face across a cut
breachprotocol profile

ByteDance's Seedance 2.5 generates a 30-second single take, and still cannot promise a face across a cut

#videogeneration #bytedance #generativevideo #productlaunch
Comments
4 min read
DeepSeek's low effort setting writes more than its high setting, because the dial is just a prompt
breachprotocol profile

DeepSeek's low effort setting writes more than its high setting, because the dial is just a prompt

#deepseek #prompting #apidesign #inference
Comments
4 min read
DistillAlign explains why fast video models get prettier and more repetitive at the same time
breachprotocol profile

DistillAlign explains why fast video models get prettier and more repetitive at the same time

#videogeneration #distillation #diffusionmodels #research
Comments
4 min read
Four agent-memory papers landed in a week, and none tested what happens when an attacker controls the writes
breachprotocol profile

Four agent-memory papers landed in a week, and none tested what happens when an attacker controls the writes

#cybersecurity #aisecurity #agentmemory #aiagents
Comments
5 min read
Kimi K3 runs in 8 gigabytes of RAM, at 33 seconds per token
breachprotocol profile

Kimi K3 runs in 8 gigabytes of RAM, at 33 seconds per token

#kimi #localai #mixtureofexperts #inference
Comments
4 min read
llama.cpp shipped DSpark for DeepSeek V4 Flash, and almost everyone called it the wrong name
breachprotocol profile

llama.cpp shipped DSpark for DeepSeek V4 Flash, and almost everyone called it the wrong name

#deepseek #llamacpp #speculativedecoding #localai
Comments
4 min read
NeurIPS rebuttal week ended with authors, reviewers and chairs all reporting the same silence
breachprotocol profile

NeurIPS rebuttal week ended with authors, reviewers and chairs all reporting the same silence

#peerreview #neurips #academia #researchculture
Comments
4 min read
Quantizing V4 Flash's KV cache in llama.cpp changes which tokens it picks
breachprotocol profile

Quantizing V4 Flash's KV cache in llama.cpp changes which tokens it picks

#deepseek #kvcache #quantization #localai
Comments
4 min read
The "2x GB200 bandwidth" Chinese chip claim is a 2027 projection, and the arithmetic gives 1.67x
breachprotocol profile

The "2x GB200 bandwidth" Chinese chip claim is a 2027 projection, and the arithmetic gives 1.67x

#hardware #accelerators #china #memorybandwidth
Comments
4 min read
A judge did not rule that ChatGPT users have no rights to their chats
breachprotocol profile

A judge did not rule that ChatGPT users have no rights to their chats

#cybersecurity #privacy #dataretention #legal
Comments
4 min read
A month after the Hugging Face breach, there is still no lawsuit
breachprotocol profile

A month after the Hugging Face breach, there is still no lawsuit

#cybersecurity #aisecurity #agentsecurity #incidentresponse
Comments
4 min read
AI financial advice works when you hand it a full financial plan
breachprotocol profile

AI financial advice works when you hand it a full financial plan

#finance #llmevaluation #bias #consumerai
Comments
4 min read
China did not give away free models. It built a governance body.
breachprotocol profile

China did not give away free models. It built a governance body.

#policy #geopolitics #openweights #china
Comments
4 min read
Figure's viral ladder climb is a two-hour stair timelapse
breachprotocol profile

Figure's viral ladder climb is a two-hour stair timelapse

#robotics #humanoids #figure #autonomy
Comments
4 min read
llama.cpp ships the fix that lets DeepSeek V4 Flash call tools mid-thought
breachprotocol profile

llama.cpp ships the fix that lets DeepSeek V4 Flash call tools mid-thought

#deepseek #localinference #openweights #tooluse
Comments
4 min read
OpenAI publishes ten mathematics claims with Lean proofs and no named authors
breachprotocol profile

OpenAI publishes ten mathematics claims with Lean proofs and no named authors

#openai #mathematics #formalverification #reasoning
Comments
4 min read
Qwen trained its phone agent on a lab of more than a hundred real phones
breachprotocol profile

Qwen trained its phone agent on a lab of more than a hundred real phones

#agents #guiagents #qwen #benchmarks
Comments
5 min read
Reddit's US daily users slipped while everything else grew
breachprotocol profile

Reddit's US daily users slipped while everything else grew

#industry #reddit #search #aisummaries
Comments
4 min read
Seven new papers cannot agree what a world model is made of
breachprotocol profile

Seven new papers cannot agree what a world model is made of

#worldmodels #videogeneration #robotics #research
Comments
4 min read
The EU AI Act's transparency rules start today. The high-risk rules do not.
breachprotocol profile

The EU AI Act's transparency rules start today. The high-risk rules do not.

#policy #euaiact #regulation #transparency
Comments
4 min read
The non-sofic group is the one OpenAI claim a computer can check
breachprotocol profile

The non-sofic group is the one OpenAI claim a computer can check

#openai #mathematics #formalverification #lean
Comments
4 min read
AskChem indexes 2.4 million individual chemistry claims instead of 147,000 papers
breachprotocol profile

AskChem indexes 2.4 million individual chemistry claims instead of 147,000 papers

#science #retrieval #agents #chemistry
Comments
3 min read
A decades-old keyword ranker beat the search agent once the document pile passed 10 million tokens
breachprotocol profile

A decades-old keyword ranker beat the search agent once the document pile passed 10 million tokens

#rag #retrieval #agents #benchmarks
Comments
3 min read
DeepSeek re-trained V4 Flash without touching the architecture and its coding-agent score went from 7 to 54
breachprotocol profile

DeepSeek re-trained V4 Flash without touching the architecture and its coding-agent score went from 7 to 54

#deepseek #openweights #agents #coding
Comments
4 min read
Training on the best of K guesses is a third scaling axis alongside parameters and data
breachprotocol profile

Training on the best of K guesses is a third scaling axis alongside parameters and data

#generativemodels #training #scalinglaws #diffusion
Comments
3 min read
An open 35B model trained to evolve its own machine-learning code nearly doubled its base model's medal rate
breachprotocol profile

An open 35B model trained to evolve its own machine-learning code nearly doubled its base model's medal rate

#recursiveselfimprovement #agents #openweights #machinelearningengineering
Comments
3 min read
Metis puts an agent's memory inside the model instead of in a database beside it
breachprotocol profile

Metis puts an agent's memory inside the model instead of in a database beside it

#agentmemory #architecture #research #longcontext
Comments
4 min read
METR published the access list an outside investigator would need to explain why an AI agent misbehaved
breachprotocol profile

METR published the access list an outside investigator would need to explain why an AI agent misbehaved

#cybersecurity #aisafety #governance #incident
Comments
4 min read
Twenty-three frontier models were handed a hacked server to clean up and none finished the job
breachprotocol profile

Twenty-three frontier models were handed a hacked server to clean up and none finished the job

#cybersecurity #aisecurity #agents #benchmarks
Comments
3 min read
13.6% of SWE-bench Verified pairs a bug report with a patch that does not match it
breachprotocol profile

13.6% of SWE-bench Verified pairs a bug report with a patch that does not match it

#benchmarks #coding #evaluation #swebench
Comments
3 min read
One planted document flipped more than half of deep-research reports to a false conclusion
breachprotocol profile

One planted document flipped more than half of deep-research reports to a false conclusion

#cybersecurity #aisecurity #datapoisoning #agents
Comments
3 min read
Letting an agent organise its own memory halved retrieval cost and improved no answers
breachprotocol profile

Letting an agent organise its own memory halved retrieval cost and improved no answers

#agentmemory #agents #evaluation #research
Comments
3 min read
Asking a model to check its own work lost every comparison against just sampling more answers
breachprotocol profile

Asking a model to check its own work lost every comparison against just sampling more answers

#reasoning #testtimecompute #evaluation #prompting
Comments
3 min read
loading...