Videos

Strata local AI runs on a single 12GB 3060 and works surprisingly well, averaging about 49 tokens/sec decode once...

Learning to build software is still worth it, but the focus shifts from writing code to 'intent engineering' –...

The best LLMs only follow complex enterprise policy instructions 36.2% of the time under strict grading, according...

Venice AI's approach to privacy in AI agents centers on keeping secrets out of the LLM reasoning layer by tying...

Agents fail at data tasks not because they can't write SQL — they fail because they lack a persistent understanding...

GraphRAG is not a replacement for vector search but an additive layer that lets an agent hop between nodes through...

Voice agents fail because the industry optimizes speed instead of behavior, argues Hall: latency is a fixed budget...

AI agents don't read policy documents — they execute API calls. Governance must be enforced at the infrastructure...

Live API calls (MCP) alone cannot handle semantic queries across large datasets; a context graph with synced data,...

The video argues that the most effective AI builders focus on 'harness engineering' — building wrappers around...

C2C (cache-to-cache) passes a sharer model's KV cache directly to a receiver model via a trained fuser, skipping...

A single loss change, swapping MSE for RMSSE in Omniquant, recovers about 63% of the accuracy gap when quantizing a...

A 300B parameter model can run on a 128GB mini PC using aggressive quantization to ~2-3 bits per weight, sparse...

A builder is running Kimi K2, a 2.8 trillion parameter model, on 16 Nvidia DGX Spark boxes at about 21 tokens per...

Gemma 4 is Google DeepMind's latest open model family, released under Apache 2.0, with sizes from 2B to 31B...

A Claude agent that teaches like Andrej Karpathy can be built by crawling his public content, compiling it into a...

Context Language Models (CLMs) give the model control over its own context, allowing it to delete, edit, and...

A community port of Splash Inco's engine roughly doubles inference speed on 2021 M1 Max Macs for the Quinn 3.8 27B...

The 'latent shopper problem' is the key challenge for enterprise AI: users often don't know what they're looking...

Claude Sonnet 5.5 absolutely demolished GPT-6 Soul in the presenter's visual task tests, including a GTA clone and a...

Weights & Biases' Arya agent uses its own production traces and offline evaluations to generate new tasks and prompt...

Anthropic's official Opus 5.5 prompt guide is built around one practical shift: the model now defaults to medium...

MiMo-V3 introduces two techniques—KV bridging and KV reuse—that shrink a million-token KV cache from 12.09 GB to...

Spotify's Large Taste Model, used daily by a quarter of US premium subscribers, shifts personalization from ranking...

Recommendation systems follow the same power-law scaling curves as LLMs. Meta's research lead argues the field is...

DoorDash treats LLMs as offline reasoning engines for search and personalization rather than serving them online....

GPT-6 Luna is OpenAI's new cost-performance workhorse: 10 cents per million input tokens, 50 cents per million...

Claude Opus 5.5 is a cheaper, capable successor to Opus 5: Anthropic claims it matches a top-tier model (called...

Anthropic's Claude Opus 5.5 matches the performance of Fable 5.1 on most tasks while costing 40-51% less to run and...

Splitting a voice agent's prompt into per-step contexts cuts token usage by more than half in production, according...

Jev is a decision-only model from Type Safe that outputs a fixed-choice answer with a confidence score instead of...

A 1998 StarCraft: Brood War benchmark called Brood War Bench shows frontier AI models lose badly in real-time...

On a 6GB GPU, the deciding factor for local coding models is not the download size but the memory left over after...

OpenAI's inference load balancer evolved from a feedback-loop-based P controller to a control plane/data plane...

Jev, a system one model from Typesafe, claims up to 200x speedup on decision tasks by selecting from predefined...

DeepSeek V4.1 Flash uses a Causal Encoder-Decoder architecture that halves prefill compute by processing input...

Reinforcement learning can make search agents roughly 20 times faster and about 100 times cheaper than frontier...

Exa is building a search engine designed specifically for AI agents, arguing that AI systems will soon perform a...

BM25, a 30-year-old lexical scoring function, is making a strong comeback in agentic search because modern LLMs can...

EliseAI's cascaded voice agent architecture uses a streaming speculative transcriber (layering fast Flux on accurate...

tldraw's spatial harness for agents turns the infinite canvas into a medium where LLMs can understand 2D space, set...

AI models do not think like humans; they lack the abductive reasoning that humans use to invent explanations from...

GPT-6 Astra beat Fable 5.1 in 10 of 15 real-world use cases, costing $186 less overall ($327 vs $513) but taking 1...

Two 8.4 GB GGUF files of Qwen 3.8 27B differ by 7.7 points on average benchmarks because one uses a new quantization...

MiniMax's M3 is the only top-five open-source model that is truly multimodal, packing a 1 million-token context,...

The looped model Hugging (3.5B parameters, 32 laps) doubles its science score by running more laps, but its silent...

GLM 5.3 Flash delivers near-flagship performance at roughly 1/9th the cost of GLM 5.3, but its heavy token usage for...

Code has won as the action space for AI agents, but nearly every harness still uses JSON-era timing by waiting for a...

Strands Agents is an open-source framework from AWS that adds an LLM-powered agent layer to robots, enabling natural...

Adobe's Agentic Sites project demonstrates that hyper-personalized web pages can be generated in about 1.1 seconds...