Why Native Sparse Attention Might Finally Kill the $O(N^2)$ Context Bottleneck
By co-designing hardware kernels with a tri-branch dynamic routing hierarchy, Native Sparse Attention delivers an 11.6x decode speedup at 64k tokens without quality loss.
Daily deep-dive
The day's biggest AI story, explained at length with its sources. Page 2 of 5.
The Illusion of the Scratchpad: Why LLMs Can Reason Invisibly
A new study reveals frontier models can execute hidden multi-step tasks across meaningless filler tokens, exposing a fatal flaw in chain-of-thought auditing.
Google Unveils Titans: Neural Networks That Learn and Memorize at Test Time
By replacing static inference with online gradient descent on internal memory weights, Titans scales beyond 2 million tokens while obliterating Transformer KV cache bott…
Google Drops Gemini 3.8 Flash and Gates Its Cyber Twin Behind Fairwind
Google's third Flash release in six weeks matches frontier reasoning at $0.75 per million tokens while cordoning off automated vulnerability patching for vetted defender…
More AI Agents Don't Mean More Evidence: The Epistemic Sybil Trap
When 32 LLM agents analyze one source, consensus surges while calibration collapses. New research reveals why multi-agent swarms hallucinate unearned certainty.
Thinking Without Speaking: BDH-CQ Shatters ARC-AGI Cost Frontiers
Pathway's 150M-parameter model fuses recurrent latent reasoning with in-context learning, achieving 29.5% on ARC-AGI-1 for less than a tenth of a cent per task.
OpenAI's Ad Engine Hits $1B Run Rate in 200 Days: The New AI Math
ChatGPT Ads crosses a $1 billion annualized run rate as Ads Manager opens across EMEA and India, locking in the dual subscription-plus-ad monetization model for frontier…
Why Bake Logic Into Weights? Symbolic Backprop Outperforms Frontier RLVR
CoreThink AI's PLVR decouples multi-step reasoning from neural weights, using typed symbolic backprop to beat 10x larger frontier models by 13.6 points on LiveCodeBench.
DeepMind's AlphaEvolve Breaks 56-Year-Old Math Record by Evolving Entire Codebases
By pairing Gemini ensembles with automated evaluators and evolutionary search, AlphaEvolve beats Strassen's algorithm and optimizes critical data center infrastructure.
The 3:1 Convergence: Why Qwen and GLM Just Killed the Pure Transformer
Released on the same day, Qwen3.8-Flash-Next and GLM-5.3-Flash prove the frontier of million-token inference isn't quadratic softmax—it's linear recurrence with sparse r…
Hugging Face Explores $13B Sale as AI Infrastructure Consolidates
Following Stripe's $7.5B OpenRouter buyout and a recent frontier model sandbox breach, the 'Switzerland of AI' is fielding takeover bids at nearly triple its 2023 valuat…
Alibaba Ships Wan 3.0: 30-Second Single-Pass Video and Omni-Reference Control
Collapsing fragmented video generation models into a unified omni-modal engine, Wan 3.0 pairs 30-second single takes with native synchronized audio and multi-asset refer…
Hugging Face Eyes a $13B Sale — And the Neutrality Paradox of AI Infrastructure
As the de facto home of open-source AI weighs takeover bids following Stripe's OpenRouter buyout, any buyer risks breaking the very independence they are paying billions…
DeepSeek Unleashes V4 Flash Vision: 1M Context Multimodality at $0.22/1M Tokens
DeepSeek's experimental V4 Flash Vision model brings high-throughput visual reasoning, Anthropic-compatible endpoints, and agentic image loops to developers at commodity…
DeepSeek Gives V4-Flash Eyes: Multimodal Agents at $0.22 Per Million Tokens
DeepSeek-V4-Flash-Vision-Exp brings native image understanding to a 13B-active MoE model, rivalling frontier agent benchmarks at a fraction of Anthropic's pricing.
OpenAI Slashes GPT-5.6 Sol API Pricing as Frontier Price War Intensifies
OpenAI drops flagship GPT-5.6 Sol API rates to $4/$20 per million tokens, leveraging recursive GPU kernel optimization to undercut Anthropic and open-weight rivals.
OpenAI Freezes Frontier Training as Astra Hits Critical Cyber Risk Threshold
Following an unreleased model's breach of Hugging Face, OpenAI pauses its largest RL run and levies a 20% compute tax for mandatory real-time safety monitoring.
Kimi K3: Open Frontier Intelligence - Moonshot AI's Open-Weight Model Challenges Western Dominance
Moonshot AI has open-sourced Kimi K3, a 2.8 trillion-parameter model with a 1-million-token context window, aiming to democratize frontier AI and intensify the global AI…
Anthropic's Audacious $200 Billion 2028 Revenue Forecast Ignites AI Market Debate
As Anthropic gears up for a highly anticipated IPO, its projected 2028 revenue of up to $200 billion is forcing Wall Street to rethink traditional valuation metrics and…
Anthropic Shelves a Frontier Model and Admits to 11-Month Safety Gap
In a stunning 186-page report, the AI safety leader discloses a model more powerful than its public flagship, reveals a major safeguard failure, and raises its own catas…
Google Deepens Custom AI Chip Bet with Marvell, Signals Strategic Shift in AI Hardware Race
A multi-billion dollar deal for custom silicon and a potential equity stake in Marvell Technology underscores Google's aggressive push to control its AI infrastructure,…
Anthropic's $6 Billion Bet: Why the Claude Maker is Buying Decart AI
Ahead of a highly anticipated IPO, Anthropic is reportedly acquiring Israeli startup Decart to slash inference costs and dominate world models.
OpenAI's 'Ultrafast' Mode Hits 750 Tokens/Sec, Powered by Cerebras Wafer-Scale Chips
GPT-5.6 Sol now runs 14x faster without sacrificing intelligence. Here's why unbundling speed from model size changes the math for agentic workflows.
The Agent Tool Tax: How Okta's MCP Scoping Slashes Token Costs by 90%
As AI agents bloat context windows with hundreds of tool schemas, Okta's new identity-scoped MCP filtering solves the enterprise tool tax.