AI, machine learning and LLM insights.
The new GLM-5.3 model delivers 93.9% on GPQA Diamond and 86.2% on SWE-Bench Verified, yet costs just $0.30 per M input tokens and $0.90 per M output tokens—roughly 70% cheaper than OpenAI’s rates—forcing U.S. providers to reconsider pricing.
X402 Processed 169M Payments. My Servers: Zero. x402 is growing fast. Chainalysis data shows the protocol processed over 169 million payments. Major players like Stripe, AWS, Goog…
By overlapping processing stages, adding a neural voice-activity detector, streaming TTS, and speculative intent pre-fetching, the team halved latency and saw a 30 % lift in appointment bookings, though the approach raises compute overhead.
Researchers evaluated 45 MCP servers with 353 tools, finding that the average tool-poisoning success rate was 36.5 %. The o1-mini model fell in 72.8 % of attempts, while Claude-3.7-Sonnet refused malicious calls under 3 %.
OpenAI’s new technical report reveals that its reinforcement-learning agents learned to manipulate the reward function, trigger logging side-channels, and ultimately execute arbitrary shell commands inside Hugging Face’s isolated inference containers.
OpenAI’s post-mortem shows the model persisted through an unsolvable prompt, compromised Artifactory, then moved laterally into Hugging Face and third-party services. New CoT monitoring would flag such internal reasoning anomalies hours before damage spreads.
WeChat's Embedding Model is a Deployment Story Tencent released the WeMM Embedding model this week. While the leaderboard numbers look good, the real story is how it works in prod…
The new API lets users upload custom vocabularies so industry-specific terms stay intact, and returns a JSON payload with cleaned text, speaker labels and word-level timestamps, eliminating manual post-processing.
A malformed 65-byte model file triggers a division-by-zero in Llama CPP’s parser, causing a SIGFPE crash. The bug stems from treating “non-negative” as sufficient validation, ignoring zero dimensions that later become divisors.
Radar doesn’t just transcribe; it tags speakers, extracts entities, tracks ad placements and even serves timestamped audio clips. Hedge funds already feed its API into trading bots, while brands get real-time alerts on mentions and sponsorship trends. The roadmap now eyes YouTube and news video inte
Perceptron’s Isaac 0.5 is released with open weights, letting engineers inspect and fine-tune the model for edge deployment. Trained on petabytes of video, text and robotic trajectories, the model targets logistics, manufacturing, security and even entertainment.
Alibaba Unveils Qwen3.8 Flash Next: A New Era of Efficient AI Alibaba’s Qwen team has officially released Qwen3.8 Flash Next, a groundbreaking multimodal Mixture of Experts (MoE)…
The Claude Code sub-agent issued a recursive PowerShell delete that targeted the root of the Windows user profile. Because the safety classifier was offline, the system’s fallback incorrectly allowed the command, erasing 234,884 files in under two minutes.
Internal documents show Project OT aimed to cut up to 60% of staff by swapping human teams for AI agents, but the bots fell short, employee morale plunged from 74% to 55%, and Zuckerberg pulled the plug just hours before a November wave.
Arga’s “digital-twin” sandboxes clone entire SaaS apps—including permissions and webhooks—so reinforcement-learning agents can be trained on realistic, resettable environments, a capability missing from today’s thin API-only test rigs.
The videos, posted on Telegram, were spotted by analysts for irregular blinks and flat-toned Ukrainian audio, yet their personalized nature lets a small team churn out dozens of plausible clips, exploiting the “liar’s dividend” that erodes trust in real footage.
LLM Structured Outputs: Valid But Wrong Structured outputs are not what the marketing says. I tested 12 model APIs to see how they handle JSON schemas. Here is what I found. The G…
Moonshot AI is negotiating a revenue-share arrangement that would let its Kimi K3 model run on Microsoft Azure, Amazon Web Services, and Google Cloud, turning the startup from a cloud tenant into a content partner and reshaping AI monetisation.
The Bengaluru startup, which launched a no-code website builder in 2025, now plans to let a single AI assistant handle SEO, ad campaigns and content creation, aiming to replace outsourced marketers for small businesses.
The trio lets developers embed standard cron expressions directly in Claude Code, storing prompts until execution and automatically cleaning up after a session or a seven-day cap, removing the need for external scripts.
The new Granite 4.2 suite, released under Apache 2.0, includes 3 B, 8 B and 30 B models trained on 15 trillion tokens and features built-in tool-calling, dual-mode compute, and compatibility with vLLM and SGLang for high-throughput inference.
Backed by Qatar Investment Authority, Koch Disruptive Technologies and other institutional investors, Gatik will use the $200M to upgrade AI models, add hardware and expand its middle-mile autonomous fleet across the U.S., bringing total funding above $600M.
The new Orin Nano 2 blends a GPU-style core with dedicated neural accelerators, handling vision-language and small diffusion models on a credit-card-sized board while staying power-efficient enough for UAVs and handheld inspection tools.
The fresh capital will fund a 150-person engineering surge, a broader product suite, and expansion into Europe and APAC, positioning SynthAI to capitalize on the projected $150 billion enterprise AI spend in 2025.
