MarkTechPost
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Perplexity releases pplx-embed: embedding models that change the rules of search
Perplexity introduced the multilingual pplx-embed family of embedding models based on Qwen3 with bidirectional attention — and claims it sets a new quality standard for search at internet scale.
Liquid AI challenges the race for bigger models: LFM2's hybrid architecture changes the rules
Startup Liquid AI unveiled a 24-billion-parameter model that combines the attention mechanism with convolutions. The goal is to address the scaling problems facing the entire large language model industry.
Composio open-sources Agent Orchestrator, a tool for multi-agent systems beyond ReAct
Startup Composio has released an open-source agent orchestrator that tackles the main problem of AI in production: the fragility of simple ReAct loops. The new approach promises scalable multi-agent workflows.
OpenAI’s WebSocket mode changes the game for voice AI
OpenAI has introduced a WebSocket mode for the Realtime API that eliminates the main enemy of voice AI agents: latency. Three separate stages of voice processing are now compressed into a single stream.
Taalas challenges GPUs: 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 specialized hardwired chips. The result: 17,000 tokens per second.
TruLens: how to stop blindly trusting LLMs and start measuring quality
Language models still remain black boxes for most developers. TruLens offers a toolkit that makes every step of an LLM workflow transparent and measurable — from input data to the final answer.
Google argues for thinking deeper, not longer — and halves inference costs
Researchers from Google and the University of Virginia have shown that a long chain of reasoning is not necessarily a good one. The new Deep-Thinking Ratio metric improves LLM accuracy while cutting compute costs in half
Google DeepMind introduced Lyria 3 for creating music from photos
Google DeepMind's new Lyria 3 model marks a breakthrough in audio generation. AI can now create full tracks with vocals and lyrics using only photos or text descriptions.
Tavus unveils Phoenix-4: video avatars with lifelike emotions and instant response
The new Phoenix-4 model overcomes the «uncanny valley» effect by adding emotional intelligence to digital avatars and cutting generation latency to a record 600 milliseconds.
Anthropic introduces Claude 4.6 Sonnet with a 1 million-token context window
Anthropic enters the era of reasoning models. Claude 4.6 Sonnet gets a 1 million-token context window and an innovative fact-verification system through code execution.
Google DeepMind introduced a framework for intelligent task delegation among AI agents
Researchers developed a system for building a safe “agentic web,” where autonomous programs can interact efficiently and scale.
Google introduced WebMCP: a new standard for AI agents interacting with the web
Google introduced the WebMCP protocol, which will allow AI agents to interact with websites directly through the code structure, bypassing slow and resource-intensive screenshot analysis.
Memory architecture: how to teach AI agents to remember what matters and reason logically
An analysis of a new memory concept for neural networks: the shift from simply storing dialogue history to structured knowledge bases for deep data analysis.
Guide to Building a Synthetic Data Pipeline with CTGAN and SDV
Learn how to build a reliable tabular data generation system that ensures high accuracy and preserves all statistical patterns for safe neural network training.
How to Train AI on Human Preferences: A Guide to DPO and QLoRA
A new tutorial shows how to efficiently tune language models to human preferences without using heavyweight reward models, with just one GPU.
Google Gemini 3 Deep Think: new record in general intelligence tests
Google unveiled Gemini 3 Deep Think. The model scored 84.6% on the ARC-AGI-2 test, bringing the industry closer to AGI through deep logical reasoning mechanisms.
OpenAI introduced GPT-5.3-Codex-Spark: a speed revolution in software development
OpenAI released an ultra-fast coding model. Thanks to Cerebras chips, it delivers 1,000 tokens per second, 15 times faster than standard versions.
How to Train Matryoshka Embeddings for Ultra-Fast Data Search
Learn how Matryoshka Representation Learning makes it possible to shrink vectors to 64 dimensions without a significant loss in search accuracy.
NVIDIA introduced KVTC: an LLM cache compression technology that will make neural networks run 20x faster
NVIDIA researchers introduced KVTC, an innovative KV-cache compression method. The technology can reduce memory consumption by 20x, addressing the main scaling problem of language models.
How to simplify neural network development: a guide to using Einops for tensor computations
The Einops library is becoming the standard for working with tensors in deep learning. We explain how to make code cleaner and avoid errors in Vision and Attention architectures.
Google introduced NAI: Gemini-based adaptive interfaces for inclusive design
Google Research proposes rethinking software accessibility. The NAI framework uses Gemini AI agents to instantly adapt the interface to the user's physical abilities and context.
Microsoft unveiled OrbitalBrain: distributed AI learning right in space
# Microsoft Presented OrbitalBrain: Artificial Intelligence Training in Space Changes the Rules Earth satellites collect petabytes of high-resolution images every day, but most of this data never…
OAT: how action tokenization brings robots closer to LLM capabilities
Robotics stands on the threshold of a new era, largely thanks to advances in large language models (LLM).
MLflow for LLM: prompt versioning and regression testing
Development and deployment of large language models (LLM) is a complex task requiring not only significant computational resources but also effective tools for management and quality control.
ByteDance Released Protenix-v1: Open-Source Model for Biomolecular Structure Prediction
ByteDance, known for its developments in artificial intelligence, has presented Protenix-v1, an ambitious open-source project aimed at reproducing the capabilities of AlphaFold3 (AF3) in the field of…
PaperBanana: Google Taught Neural Networks to Generate Publication-Quality Charts
You've ever seen a PhD student cry before a Nature or Science deadline? Usually it happens at three in the morning, when the Matplotlib package once again refuses to align the chart legend, and the…
Waymo World Model: Google's autonomous vehicles now train in their own 'Matrix
Autonomous driving has always hit the same wall: reality is predictable in 99% of cases and frighteningly chaotic in the remaining one percent.
Claude Opus 4.6: Anthropic teaches neural networks to think slowly and work for you
While the industry stands frozen waiting for OpenAI's next major move, Anthropic continues methodically capturing territory in professional AI adoption.
GPT-5.3-Codex: OpenAI Transforms Coding Into an Autonomous Process (and Does It Faster)
Remember when Copilot seemed like the pinnacle of technological progress? OpenAI has just proven that was only the beginning.
Voxtral Transcribe 2: Mistral reminded us why we still need European neural networks
While everyone was waiting for Mistral to release another iteration of a large language model, the French decided to come in from the side and strike at the speech recognition market.