In Brief
VC isn't VC anymore – understanding the rise of Cancer Capital
Why this matters now: Anil Dash's argument that modern mega‑firms act as "Cancer Capital" matters because founders, employees, and public investors face different incentives and exit dynamics than they did under traditional venture models.
Anil Dash lays out a forceful critique: a small number of massive investment firms have blended private equity and venture playbooks, eroded previous accountability fences, and now wield outsized, self‑reinforcing influence over startups and markets, according to Dash's post. He writes bluntly that "it’s not 'venture capital' anymore," and warns that insiders can offload risky paper onto public and retirement investors while avoiding the discipline of truly public markets.
"IPOs are increasingly a way for insiders to 'dump the paper on the retail' rather than true financing or public‑market discovery."
Why that will keep bubbling up: founders lose leverage, IPOs can become liquidity events for insiders instead of validation, and political spending by big firms changes the regulatory backdrop. Treat Dash's piece as a clarifying frame rather than fresh empirical proof — it leans on experience and anecdotes — but it's a lens that explains a lot of current deal behavior and why regulators and founders are suddenly paying attention.
Static Allocation, Constant Work
Why this matters now: Matklad's memory‑safety manifesto is immediately useful for engineers building ultra‑low‑latency systems where occasional allocator bugs become production catastrophes.
Matklad revisits an old reliability trick with modern urgency in his post: allocate typed, fixed‑size pools up front, reject or reserve excess slots, and design for "constant work" so your hot path never has to chase frees, reallocations, or complex generation tracking. The pattern reduces whole classes of use‑after‑free bugs and improves cache locality and predictability.
"Static allocation gives you peace of mind."
The tradeoff is explicit: you spend memory and accept capacity limits in exchange for far simpler correctness and latency behavior. For trading engines or real‑time systems, that's often a good trade.
Deep Dive
K2 Horizon: Frontier Performance, Radically Open
Why this matters now: IFM's K2 Horizon release matters because it pairs a full, reproducible research stack with new inference and architecture ideas — if the claims hold, that changes what open models can do across devices and scales.
IFM (Institute for Foundation Models) published K2 Horizon as a connected fleet of six models from 0.9B parameters up to a sparse 375B-A23B, and pitched the release as "Frontier Performance, Radically Open" in their announcement. The release includes not just weights but intermediate checkpoints, training code, data recipes, and fine‑grained logs — a rare level of transparency for frontier models.
"K2 Horizon is also our most comprehensive open release to date."
Two technical pieces deserve quick, plain explanations. First, MoVA (Mixture‑of‑Value Attention) is IFM's sparse‑attention trick: instead of attending everywhere, the model routes attention selectively to relevant heads or experts so large sparse models can scale without a quadratic attention cost. Second, Uno is a "lossless" diffusion‑distillation adapter that speeds up inference by converting slower sampling steps into a distilled, deterministic adaptor — the claim is faster responses with no quality drop.
IFM also touts cross‑scale competitiveness: SOTA claims at 0.9B/3.7B/7B and "strong" results for the 32B and sparse 375B models. Community reaction on Hacker News and the repo comments split into two threads: many celebrate another fully open stack — important for reproducibility and research — while others flagged selective benchmarking and missing head‑to‑head comparisons (for example, the dense 32B wasn't directly compared to some contemporaries). A few users reported practical wrinkles at launch: login friction and some empty repos.
What to watch next:
- Independent reproductions of the benchmarks using the published checkpoints and logs. IFM made that easy by design, but reproducibility still depends on clear measurement scripts and unambiguous baselines.
- Whether MoVA and Uno hold up across tasks and inference runtimes, particularly on edge devices where latency and memory matter.
- How the ecosystem uses the intermediate checkpoints and branches IFM provided — this is where openness turns into community knowledge about capability emergence.
Bottom line: K2 Horizon is a big, transparent bet — one that invites scrutiny. If the claims are validated, the release could accelerate robust, open research into sparse scaling and faster inference across device classes.
Closing Thought
Three different kinds of friction showed up in today's reading: friction to replicate scientific claims (and IFM's response by opening everything), friction in markets and governance as capital concentrates (Anil Dash's "Cancer Capital"), and friction in code where dynamic allocation hides catastrophic bugs (Matklad's static allocation pattern). Each piece offers a practical answer: publish the pipeline, rethink accountability and incentives, and constrain allocation in latency‑critical paths. Watch for independent K2 reproductions, more regulatory noise around mega‑funds, and the growing adoption of static allocation patterns where predictability matters.