⚡️ Claude Code's 50% usage boost ends August 19

Anthropic's been running a quiet promo since May 13 that bumped Claude Code weekly limits by 50%. It ends Tuesday, and limits revert to normal.

Generous while it lasted. Time to pace yourself.

⚡️ Cerebras CS-4: 4,400 tokens/sec and up to 30x faster than GPUs

Three WSE-3 Turbo wafers, 750 PFLOPS, 129.6 PB/s memory bandwidth. On GPT-OSS-120B it hits over 4,400 tokens/sec per user, vs. GPU configs it's not even close.

The WSE-3 Turbo doubles AI compute to 250 PFLOPS per wafer and doubles memory bandwidth to 43.2 PB/s. Same 5nm die, just pushed harder.

Specialized silicon keeps quietly eating the GPU lunch.

⚡️ A real fly brain is now living on your Mac desktop

This open-source project runs a live spiking simulation of the actual FlyWire connectome (138k neurons) and renders a 3D fly that walks across your windows, grooms, sleeps, and flees your cursor using the same neural wiring a real fruit fly has.

Clones and runs in under 30 seconds. Wild use of real neuroscience data.

⚡️ 10M vectors, 4GB RAM. TurboVec makes vector search embarrassingly cheap

Rust implementation of Google Research's TurboQuant lands on GitHub with Python bindings. Float32 storage for 10M docs eats 31GB. Turbovec fits it in 4GB and searches faster than FAISS.

No separate training phase, online ingest, add vectors as you go. Local RAG that doesn't need a server rack is suddenly very real.

🔥 Mojo goes full open source. Finally.

Chris Lattner (LLVM, Swift, now Qualcomm) just dropped the Mojo compiler under Apache 2.0. The standard lib was already open. Now it's all there.

Python syntax. Systems-level speed. GPU-native. And yes, the closed compiler was probably the thing quietly killing adoption this whole time.

⚡️ Ornith-1.5 teaches itself to improve and it's open weights

Post-trained on Qwen, Ornith-1.5 closes the loop: the model proposes its own tasks, builds its own scaffolds, then trains on the results. No human-curated datasets needed.

Three sizes: 9B dense, 35B MoE, 397B MoE. The 35B MoE runs fast on consumer hardware and matches Qwen3 27B at higher quant. The 397B trades blows with Claude Opus 4.8 on coding evals.

Self-improvement loops in open weights. Quietly a big deal.

🤖 Terry Tao just launched Palomar, a registry of Lean-verified math proofs

Think arXiv, but every submission gets its proof machine-checked by Lean. No more quietly broken theorems.

It's also got an LLM-backed vacuity gate, so AI-slop proofs don't sneak through.

Huge for AI theorem proving. Verified math at scale is the infrastructure that agents actually need. Source

⚡️ Stripe is buying OpenRouter for $7B+

One API key, 400+ models, 8M developers. Stripe just acquired the whole thing.

Tokens are basically currency. Stripe routes money. Routing AI calls isn't that different, and OpenRouter's 25 trillion weekly tokens is a distribution moat Stripe couldn't build from scratch.

Wild price for a 10% margin business. But the network is real.

🤖 Claude Opus 5 is verbose and users are fed up

Phrases like "load-bearing" now litter codebases. Outputs balloon into 8-paragraph walls when one sentence would do. Users are begging for 4.6 back.

Workarounds are spreading: "ELI5", steering files, plain-language prompts just to get a straight answer from a flagship model.

Source

⚡️ Google ditched public Git tags for Android source. Now you fill out a Google Form and wait.

GrapheneOS flagged that Google replaced Git tag pushes with Drive downloads gated behind a request form, and says responses have gotten very slow.

GPL violation? Probably not. The license lets you charge for shipping a tape. This is just the tape being a Google Form now.

Bad look either way.

⚡️ Unsloth Dynamic 3.0 squeezes >10% better accuracy from the same GGUF size

Unsloth just dropped Dynamic v3.0 quants for Qwen3-27B, with smarter per-layer quantization and a higher-quality imatrix calibration dataset tuned for coding, chat, and multilingual.

Same file size. More of the original model's brain intact.

For the ultra-low crowd (IQ2_XXS and below), MTP is stripped to save ~500MB. Tight 16GB box? They recommend UD-IQ3_XXS at 10.9GB instead.

🧠 Researchers can steal "hidden" reasoning traces from o1, Claude, and Gemini APIs

Labs encrypt chain-of-thought and send it back to clients rather than storing it server-side. Turns out those encrypted blocks are fully interchangeable across sessions, users, and models within the same provider.

Inject a trace into a weaker, less-guarded model and it just... reads it back to you. Verbatim.

The whole "hidden reasoning = IP moat" thing is looking shakier by the day.

🧠 Terence Tao just wrote the paper on what AI actually means for math

ICM 2026 keynote, now on arxiv. He doesn't debate whether AI can do research-level math. He assumes it can, then asks what mathematicians do next.

His sharpest line is that AI proofs "dwell on trivialities while obscuring the most interesting parts." A proof no human can explain should be considered incomplete, even if formally verified.

Source

🤖 Anthropic just flipped on watermarks for Claude

The technique hides a signal inside token sampling itself. No invisible ink, no metadata. When Claude picks between "overcast" and "grey" (both equally valid), a secret key steers that choice, leaving a detectable pattern across hundreds of such micro-decisions.

It's built on DeepMind's SynthID-Text approach. EU's AI Act pushed the deadline to Aug 2. Heavy edits can strip it. Detection requires the key.

Clever. And now industry-wide.

🎹 GitHub Copilot, but for your piano

A dev trained a 125M transformer to autocomplete MIDI performances on-device. Play a few notes, it continues. Runs at ~108 notes/sec on an iPhone 15.

Entirely free to try. Source

🤖 Fight Claude's word salad with another LLM

"Vomit" pipes Claude 5's bloated output through a local model to strip the pseudo-epiphanies, self-praise, and weird roundabout reasoning.

Fully local, no telemetry. The local LLM only sees what Claude tries to communicate, so your context stays private.

Source

⚡️ Google's DiffusionGemma generates text like an image diffuser. And it's fast.

DiffusionGemma is Google DeepMind's open-weight text diffusion model built on a clever trick: it's a 25.2B-parameter MoE with only ~3.8B active params per forward pass.

No retraining from scratch. They converted an existing Gemma 4 checkpoint. One tester hit ~670 tok/s on consumer hardware.

Paper

⚡️ Aaron Swartz faced 35 years. Meta scrapes the whole web for AI and gets a shrug.

Swartz was prosecuted into a corner for downloading academic papers. Meta scrapes at planetary scale to train its models, with no criminal exposure in sight.

Same act. Different consequences. Turns out the CFAA bites hard when you're 26 and broke, and barely grazes you when you're a trillion-dollar company with aligned government interests.

⚡️ Tired of babysitting coding agents? This dev built a pseudocode editor instead.

Write rough pseudocode, hit save, get real code. Intent stays stored alongside the output so the "why" doesn't vanish.

It's a PoC, but the instinct is right. Full-sentence prompting gets old fast, and agents start losing the plot on big codebases.

⚡️ Amazon buys rare books, scans them, destroys them. That's "fair use" now.

A 404 Media investigation confirmed it with a $29 AirTag: bulk book orders go to Amazon's LAS8 warehouse, where spines get cut and pages scanned for AI training data.

Meanwhile, the Internet Archive is being sued into the ground for digitizing books to preserve them.

Same act. Very different courtrooms.