posts · 2026-08-27

This one's worth opening because Linum gets specific about the grunt work of data filtering. They started in 2024 running everything on CPUs to save money—PySceneDetect for shot cuts, EAST for OCR, H.264 motion vectors to kill static clips, Haar cascades to downsample talking heads. By early 2025 they'd moved to GPU clusters with fine-tuned LLMs for classification, and by late 2025 they're using Qwen-2.5-VL-3B with RLVR for 1–4 aesthetic scoring. Two things I'd flag. First, they still use WAFT optical flow to catch long-tail low-motion clips, which tells you traditional CV signals still matter at the edges even when you've got LLM scorers. Second, the bit about iterative labeling and self-consistency being harder than expected—anyone who's tried using LLMs as classifiers knows that label drift is real. The post doesn't disclose v3 release date or model size, so treat this as an engineering status report, not a product teaser. If you're training your own video model, this is more useful than most papers.

This one's worth a look because LAION went big on video: 1.3B URLs from CommonCrawl, 80M videos downloaded, 10M total hours. They used scene detection to create clips, generated synthetic captions for both video and audio, and the resulting ViCLIP model beats InternVid by up to 2.1% on video-text benchmarks. I'd discount the margin a bit — it's not a massive leap, and the captions are synthetic, not human-labeled. But the real story here is scale. This is the largest open video corpus available, directly countering the trend where big video datasets sit locked inside proprietary companies. They also extracted 300M frames for image-text training and claim the visual distribution differs from standard web images, so it can complement existing image datasets. That's a practical bonus: one dataset feeding three modalities. The limits are clear: research-only, non-commercial, and the team flags copyright and bias concerns upfront. If you're doing multimodal pretraining, it's worth downloading and testing, but don't plan on building a commercial model on top of it.

The numbers are what make this worth clicking: a one-year-old company run by 23-year-old Noah Shinn just closed a $250M Series B co-led by Benchmark and Index Ventures, at a $2.5B valuation. The product is an AI assistant that handles errands — planning road trips, buying groceries, canceling subscriptions — and it's still in private beta. I'd discount the hype a bit. The TechCrunch piece flags privacy concerns around its broad permissions and terms of use, but doesn't spell out exactly what's problematic. An assistant that takes over your daily tasks needs access to your email, calendar, and payment info — at that permission level, the security audit and data handling design matter way more than the funding round. What's missing: which tasks it can reliably complete, what the failure rate looks like, and how the privacy architecture actually works. Until those are public, $2.5B reads like a bet on an unverified use case.

This post does the math on AI model companies in a way that clicks: buy electricity wholesale, sell it back as intelligence. Dylan Patel dropped the key numbers on Dwarkesh—Anthropic's compute costs $10–15M per megawatt, but it pulls in $50M in revenue. That inference margin then funds the next training run. In 2026, Anthropic booked its first profitable quarter: $10.9B revenue, $559M operating profit. The 5% operating margin is thin compared to the 70–80% gross margin the per-megawatt math suggests—training runs and headcount eat the rest. Tunguz also points to GLM-5.3-Flash: 18B active parameters, matches Claude Opus 4.8 on the Artificial Analysis Intelligence Index, at 90–97% lower cost. Efficiency directly boosts profit per megawatt. Nvidia's $6B Poolside acquisition plus $1B investment bets on turning model building into an industrial process—52 days from kickoff to release for Laguna S 2.1. One caveat: the Anthropic margin figures come from Patel's podcast and The Information, not an official filing. A 5% operating margin means they're not printing money yet, but the inference margin turning positive is the real signal here.

This is worth reading because it explains a real engineering tradeoff you normally have to guess at: why Cursor's Grok Bot doesn't give the model all tool definitions upfront, instead forcing it to call GetMcpTools first. The main reason isn't tool clutter confusing the model—it's KV cache economics. When tool definitions sit in the API tools parameter, any change invalidates the entire prefix cache, multiplying costs by 10x. Cursor writes the 9 dynamic tool schemas into conversation content instead, keeping only 18 static core tools plus two stable entry points in the tools array. The tool surface stays fixed, the cache discount survives. Manus independently converged on the same constraint: the serialized tool surface must remain stable; dynamism gets pushed elsewhere. The second half of the piece separates knowledge-layer loading (skills) from capability-layer loading (dynamic tools) as orthogonal dimensions—a split that only became possible after Anthropic shipped Agent Skills in October 2025, giving the knowledge layer its own independent carrier. What's missing: any official response from Cursor. Anysphere pulled the 0.18.0 installer but issued no statement.

This is worth opening because it turns a "best practice" into verifiable code evidence. The community reverse-engineered Cursor's desktop agent Grok Bot 0.18.0 and found it uses `compactionEpoch` to freeze the memory and profile sections of the system prompt, keeping them byte-identical within an epoch. The goal is dead simple: preserve KV cache prefix hits. Anthropic charges $0.30 per million cached input tokens vs. $3.00 uncached, and changing the prefix invalidates the entire cache. Agent runs have roughly a 100:1 input-to-output token ratio, so the cost lives in input—prefix stability directly determines whether the thing is affordable to run. Manus's July 2025 Context Engineering post independently reached the same causal chain: "Keep your prompt prefix stable." Two teams, a year apart, same solution—this constraint is forced by the underlying economics, not anyone's design taste. Grok Bot does two more things: injects a runtime `mcp_status` block reflecting this round's MCP failure, and spills tool definitions over 12KB to the filesystem, leaving only a path in context and using `hasReadPath` to verify the model actually read the file. Manus gave the same spillover discipline: the filesystem is the ultimate context, and compression must be restorable. Manus adds three more mechanisms not visible in Grok Bot's source: reciting goals at the context tail to fight lost-in-the-middle, keeping error traces so the model can adapt, and injecting structured variation. Only Manus's evidence chain exists for these, but the logic holds. I'd discount slightly: this is reverse-engineered code, not official docs, but code is more honest than blog posts. If you're building an agent harness, these three disciplines—freeze stable prefixes, inject runtime state, spill oversized content to disk—are directly copyable.