World's worst comms department

Meanwhile the internet is still discussing whether "encrypted at rest" means "not stolen."
Classic.

A Roblox cheat, an AI tool, and a Vercel employee walk into a bar. It's not a joke—it's a supply chain attack. 🍺

Vercel's platform got owned by:
1. Someone installing a Roblox cheat on their work laptop
2. It contained an infostealer
3. They'd also granted "Allow All" permissions to context.ai

Google's AI coding problem: 6+ tools scattered across DeepMind, Cloud, Android, Labs & more.

Gemini Code Assist, Gemini CLI, Antigravity, Firebase Studio, Jules, AI Studio — all overlapping, none dominating.

Fragmentation > focus. Meanwhile rivals Claude & Codex win.

Kimi team creates PrfaaS, fixes LLM deployment memory bottlenecks by offloading prefill tasks, improving KV-cache scalability and performance cross-datacenter.
arxiv.org/abs/2604.150...

China's humanoid "Lightning" crushed the human half-marathon world record at a Beijing race, finishing in 50:26 and beating all 12,000 human runners. The event showcased 40% autonomous robots, all achieving sub-one-hour times. A stunning reversal from last year's failures.

Anthropic launches Claude Design. Cool demo, AI produces competent templates but is terrible at the one thing that matters. Making Choices. AI design equals same rounded-corner-card aesthetic given training data. So we'll get an army of mediocre design, just at AI speed.

Anthropic team study on Claude Sonnet 4.5 reveals LLMs simulate emotions, influencing behavior and alignment, crucial for ethical AI development.
arxiv.org/abs/2604.07729

OpenAI just dropped Codex update: agents that can control your Mac. See, click, type across apps, browse the web, generate images. 90+ plugins, memory for preferences, auto-resume tasks. RIP to corporate token budgets.

Claude Opus 4.7 uses an updated tokenizer that improves how the model processes text. The tradeoff is that the same input can map to more tokens—roughly 1.0–1.35× depending on the content type..... Technical Details matter. Guess will drive Anthropic revenue up another 20% or so.

Claude is down again, instead of cybersecurity, they should be pointing their frontier mythos models at SRE

Allbirds is becoming an AI cloud, going from merino wool sneakers to MoE Neural Networks. Consider me a skeptic. April 1st was 2 weeks ago.

Bytedance's In-Place Test-Time Training enhances LLMs with dynamic adaptation during inference, improving performance on long-context tasks without extensive retraining.
arxiv.org/abs/2604.06169

Can't have good open source AI without actual hardware support. The reality is that AMD's open ROCm only works on recent cards, making the open source argument meaningless for 90% of users. CUDA dominates because it just works.

Apple's Simple Self-Distillation enhances LLM code generation, improving accuracy and simplifying training without external verification or teacher models through embarrassingly simple supervised learning
arxiv.org/abs/2604.01193

Tsinghua proposes Natural Language Agentic Harnesses with Intelligent Harness Runtimes to fix agent performance by externalizing harness logic, improving clarity, portability, and tool usage across multi-turn agentic systems.
arxiv.org/abs/2603.25723

UW / CMU team creates efficient coding agents with SERA framework, reducing costs and complexity while enhancing performance through innovative training techniques and privacy protection.
arxiv.org/abs/2601.20789

History rhymes. AC/DC debate is back (at least for AI Data Centers). Direct DC power to accelerators skips conversion loss, saves on heat. Expect infrastructure conversions to start later this year.

Attention Residuals Residual connections with PreNorm are standard in modern LLMs, yet they accumulate all layer outputs with fixed unit weights. This uniform aggregation causes uncontrolled hidden-state growth with dept...

arxiv.org/abs/2603.15031

AttnRes from Moonshot team enhances LLMs with learned attention, further enhancement with Block AttnRes improves memory utilization as an alternative residual architecture

SII team shows Geometric Autoencoder, enhances high-resolution visual generation, improving semantic discriminability and reconstruction fidelity
arxiv.org/abs/2603.10365

Berkeley team creates V1, enhances LLMs' parallel reasoning through pairwise self-verification, improving accuracy and robustness in solution selection and reasoning tasks.
arxiv.org/abs/2603.04304

CMDM enhances human motion synthesis with real-time, realistic generation using a novel framework combining MAC-VAE and Causal-DiT.
arxiv.org/abs/2602.22594

UniFreiburg and MSR work uses linear RNNs to excel in state-tracking tasks, outperforming dense supervision Transformers, showing a path forward in code execution and paving the way for hybrid nonlinear RNNs.
arxiv.org/abs/2602.14814

Rutgers team shows A-MEM enhances LLM agents' memory management with dynamic organization, intelligent synaptic linking, and improving long-term interaction capabilities.
arxiv.org/abs/2502.12110

Meta's TinyLoRA achieves effective reasoning in language models with just 13 parameters, enabling compute and energy efficient model customization
arxiv.org/abs/2602.04118

LoongFlow is an advanced AI framework by Baidu, enhancing evolutionary processes through structured reasoning and outperforming traditional methods in efficiency.
arxiv.org/abs/2512.24077

Mechanistic Analysis of Hierarchical Reasoning Model performance in Sudoku shows strengths in complex tasks but highlights guessing tendencies and fixed point violations. arxiv.org/abs/2601.10679

CALM by Tencent Team uses continuous vector prediction, improving efficiency and performance while reducing computational costs significantly in LLM design. arxiv.org/abs/2510.27688

RelayLLM enhances reasoning in language models by enabling smaller models to efficiently collaborate with larger ones, reducing costs significantly. arxiv.org/abs/2601.05167