Open Source - Page 5 of 18
Open-source AI projects, community innovations, collaborative development, and freely accessible AI tools and frameworks.
352 articles
View complete article list
IBM launches Granite Speech 4.1 2B models, hits 1.33 WER on LibriSpeech clean
An autoregressive model that translates six languages. A non-autoregressive sibling that drops translation and Japanese entirely to shave latency. Both achieve a 1.33 Word Error Rate on LibriSpeech clean.
Goodfire’s tool, a MIT Tech Review 2026 breakthrough, helps debug LLM models
Large language models are opaque, famously so. They function, often brilliantly. Yet how they actually work remains a profound mystery. The standard industry approach? Cross your fingers and filter the bad outputs.
OpenAI releases Symphony, an open-source spec to let agents self‑manage tasks
Managing OpenAI's Codex AI was a lesson in human limits. Engineers found that manually assigning tasks to more than three to five concurrent agent sessions destroyed their own focus.
Calculator Finds Prompt Caching and Lazy-Loading Save Tokens, Gains Small
Every token you feed an LLM carries a hidden tax. Before the model generates a single word, it must first process your prompt, a step called prefill.
OpenAI releases 1.5B‑parameter privacy filter with 50M params, 128‑expert MoE
Sixty times fewer parameters actually doing work. That is the headline math behind OpenAI’s newly released privacy filter: a 1.5-billion-parameter model that activates only 50 million per token.
Meta FAIR releases NeuralSet: for fMRI, M/EEG, spikes, HuggingFace embeddings
Neuroscientists and AI researchers alike have long faced a brutal bottleneck: raw neural data is huge, messy, and painfully slow to load. Meta FAIR’s new open-source package, NeuralSet, takes a radically different path.
Moonshot AI releases FlashKDA with CUTLASS kernels, batching, H2O benchmarks
The race to scale long-context AI has a new contender. Moonshot AI just dropped FlashKDA, a production-grade CUDA kernel that turns their Kimi Delta Attention (KDA) from theory into hardware reality.
AI engineers abandon LangChain for native agents amid hidden complexity
LangChain promised to save you time. It mostly wastes it. Engineers are deleting the framework from their projects, trading its neat abstractions for the raw, observable guts of native agents. The initial deal seems reasonable.
Chinese AI startups drop offshore registration amid regulator IPO warning
China's AI companies are scrambling to get back home. The offshore corporate shells that let them tap foreign cash while dodging domestic rules are no longer safe. Beijing just made that clear. The signal came from the securities regulator.
OpenMOSS releases MOSS‑Audio, encoding raw audio at 12.5 Hz for speech and music
Audio is a flood, not a trickle. Every second of speech or music carries dense layers of meaning, pitch, rhythm, timbre, transient pops, background texture, all entangled in raw waveforms.
Open-source orchestration spec 'Symphony' invites agents to build implementations
Symphony began as a single Codex session running in tmux, polling Linear and spawning sub-agents for new tasks. From that scrappy seed, its second iteration embedded itself into our main project repository, a repo already built with agents in mind.
Fast AI Power‑Use Estimator Aims to Prompt Developers, Operators to Cut Energy
AI models are power hogs. Everyone knows it. Yet almost no one pauses their race for accuracy or uptime to actually measure the drain.
OpenAI, Microsoft, Zoox Spend USD 813‑USD 1,622 on San Francisco Police Protection
The numbers are absurdly small. OpenAI paid $813.43. Microsoft’s bill came to $1,622.16. Zoox ran up a tab of $838.43. For companies worth billions, these sums are pocket change.
Meta AI releases Sapiens2, a model for pose, segmentation and albedo
Most models see a person as a shape. Meta's new Sapiens2 tries to see them as matter. Computer vision has a nasty habit of confusing appearance with identity.
DeepSeek AI unveils DeepSeek‑V4 with compressed attention for 1 M‑token contexts
The AI arms race has always been a battle of scale, until now. DeepSeek’s latest release, DeepSeek-V4, doesn’t just push context windows to a million tokens; it fundamentally rewrites the efficiency equation.
DeepSeek‑V4‑Pro‑Max tops open models, nears closed results at 1/6 Opus 4.7 cost
The fight for open models has ended. DeepSeek won. Its new V4-Pro-Max isn't just the best download you can find; it's now so close to the expensive, closed products from OpenAI and Anthropic that their pricing seems absurd.
Update: Usage Limits Draining Faster Linked to Two Unrelated Experiments
Bugs are supposed to be dumb. This one was clever, which is infinitely worse. It started with user reports: usage limits were draining faster than expected. The data proved them right.
Agent Improvement Loop Starts with Trace, Enabling Deterministic, Low‑Cost Validation
An agent is only as reliable as the feedback loop that sharpens it. In production, the difference between a competent system and a brittle one often comes down to how you validate, and more importantly, *what* you choose to validate.
Anthropic's Mythos Leak Precedes Bland AI's Norm Voice Agent Builder
The silence around Anthropic’s Mythos was always strategic, a tightly guarded model that promised to redefine the frontier. Now that silence has shattered.
Trump 'saved' women from execution—AI‑fabricated; account hit Lee Jae‑myung
The same account that nearly derailed a South Korean president has now conjured a digital miracle: Donald Trump swooping in to save eight Iranian women from execution. Except the women are AI-generated.
Featured Resources & Reviews
No Code MBA Course Review
Learn to build AI-powered apps without coding. Our comprehensive review of No Code MBA's course.
AI Tools & Resources
Curated collection of AI tools, courses, and frameworks to accelerate your AI journey.
Weekly AI Digest
Get the week's most important AI news delivered to your inbox every week.