MarkTechPost
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Perplexity releases pplx-embed, embedding models aimed at web-scale search
Perplexity has unveiled pplx-embed, a family of multilingual embedding models built on Qwen3 with bidirectional attention, and claims it sets a new quality standard for search across the entire web.
Liquid AI challenges the size race: LFM2's hybrid architecture aims to change the rules
Startup Liquid AI has unveiled a 24-billion-parameter model that combines attention with convolutions, aiming at the scaling problems facing the entire large language model industry.
Composio open-sources Agent Orchestrator, a multi-agent tool that goes beyond ReAct
Startup Composio has open-sourced an agent orchestrator aimed at the core problem of production AI: the brittleness of simple ReAct loops. It promises scalable multi-agent workflows.
OpenAI's WebSocket Mode Rewrites the Rules for Voice AI
OpenAI has introduced a WebSocket mode for its Realtime API that targets the biggest enemy of voice AI agents: latency. Three separate voice-processing stages now collapse into a single stream.
Taalas takes on the GPU: hardwired logic over flexibility for 17,000 tokens per second
Canadian startup Taalas is betting on what the AI chip industry considered heresy: replacing programmable GPUs with hardwired chips. The result: 17,000 tokens per second.
TruLens: Stop Trusting LLMs Blindly and Start Measuring Quality
For most developers, language models are still black boxes. TruLens offers tooling that makes every step of an LLM's work transparent and measurable, from input to final answer.
Google says think deeper, not longer, and cuts inference costs by half
Researchers at Google and the University of Virginia show that a longer chain of reasoning is not a better one. A new metric, Deep-Thinking Ratio, raises LLM accuracy and halves compute costs.
Google DeepMind unveils Lyria 3, a model that makes music from photos
Google DeepMind's new Lyria 3 model advances AI audio generation. It can produce full tracks with vocals and lyrics from nothing more than a photo or a text description.
Tavus unveils Phoenix-4: video avatars with lifelike emotions and instant response
The new Phoenix-4 model tackles the uncanny valley by giving digital avatars emotional intelligence and cutting generation latency to a record 600 milliseconds.
Anthropic unveils Claude 4.6 Sonnet with a 1-million-token context window
Anthropic enters the era of reasoning models. Claude 4.6 Sonnet gets a 1M-token context window and a new fact-checking system that runs code to verify answers.
Google DeepMind unveils framework for intelligent task delegation among AI agents
Researchers have designed a system for building a secure "agentic web" in which autonomous programs can interact effectively and scale.
Google unveils WebMCP, a new standard for AI agents to interact with the web
Google has introduced WebMCP, a protocol that lets AI agents work with websites directly through their code structure, skipping slow, resource-hungry screenshot analysis.
Memory Architecture: Teaching AI Agents to Remember What Matters and Reason Logically
A look at a new memory concept for neural networks: moving from simply storing dialogue history to structured knowledge bases that support deeper data analysis.
A Guide to Building a Synthetic Data Pipeline with CTGAN and SDV
Learn how to build a reliable tabular data generation system that delivers high fidelity and preserves statistical patterns, so neural networks can be trained safely.
How to Train AI on Human Preferences: A Guide to DPO and QLoRA
A new tutorial shows how to tune language models to human preferences efficiently, without heavyweight reward models, using just one GPU.
Google Gemini 3 Deep Think Sets a New Record on a General-Intelligence Benchmark
Google has unveiled Gemini 3 Deep Think. The model scored 84.6% on ARC-AGI-2, using deep reasoning mechanisms that bring the industry closer to strong AI.
OpenAI unveils GPT-5.3-Codex-Spark: a speed leap for AI-assisted coding
OpenAI has released an ultra-fast coding model. Running on Cerebras chips, it produces 1,000 tokens per second, 15 times faster than standard versions.
How to Train Matryoshka Embeddings for Ultra-Fast Data Search
Matryoshka Representation Learning can shrink vectors to 64 dimensions without a significant loss in search accuracy. Here's how it works.
NVIDIA unveils KVTC: LLM cache compression that promises 20x savings and faster neural networks
NVIDIA researchers have introduced KVTC, a new method for compressing the KV cache. It cuts memory use by 20x, tackling a core obstacle to scaling language models.
How to Simplify Neural Network Development: A Guide to Using Einops for Tensor Operations
Einops is becoming a standard tool for tensor work in deep learning. Here is how to write cleaner code and avoid errors in Vision and Attention architectures.
Google unveils NAI: Gemini-powered adaptive interfaces for inclusive design
Google Research wants to rethink software accessibility. The NAI framework uses Gemini AI agents to adapt an interface instantly to a user's physical abilities and context.
Microsoft unveils OrbitalBrain: distributed AI training directly in space
Microsoft researchers propose OrbitalBrain, an architecture for training neural networks on satellite constellations to get around slow data downlinks to Earth.
OAT: How Action Tokenization Brings Robots Closer to LLM Capabilities
A new method called OAT helps robots plan actions more efficiently using approaches similar to large language models, opening the way to scaling and flexible robot control.
MLflow for LLMs: Prompt Versioning and Regression Testing
How do you keep large language models stable? MLflow offers a way to version prompts and automate regression testing. Details in the article.
ByteDance releases Protenix-v1, an open-source model for predicting biomolecular structures
ByteDance has published Protenix-v1, an open-source model for predicting protein, DNA, RNA and ligand structures at AlphaFold3 level. Code and weights are available under Apache 2.0.
PaperBanana: Google teaches AI to draw charts that won't embarrass you in front of reviewers
AI already writes code and papers, but clean diagrams were still manual drudgery. Google's PaperBanana framework automates scientific illustrations and charts.
Waymo World Model: Google's Self-Driving Cars Now Train in Their Own "Matrix"
Waymo has launched a generative model built on Genie 3. Its self-driving cars now learn to drive in photorealistic worlds the neural network generates on the fly. No more dull test tracks.
Claude Opus 4.6: Anthropic Teaches AI to Think Slowly and Do the Work for You
Anthropic has shipped Opus 4.6. While rivals chase hype, Dario Amodei is delivering a million-token context window and agents that actually write code.
GPT-5.3-Codex: OpenAI Turns Coding Into an Autonomous Process (and Makes It Faster)
OpenAI has merged the power of GPT-5.2 with coding skills in its new 5.3-Codex model. It runs 25% faster and aims to be a full autonomous agent for your computer.
Voxtral Transcribe 2: Mistral Reminds Us Why Europe Still Needs Its Own Neural Nets
Mistral has released Voxtral Transcribe 2, a family of speech recognition models. Diarization and real-time ASR now work at production scale without extra cost.