In Brief
tmp.0ut Volume 5
Why this matters now: tmp.0ut Volume 5 revives deep systems craft with fresh writeups and interviews that teach low-level skills useful for security, reverse engineering, and compact tooling today.
The latest issue of the hacker zine leans hard into minimalism and craftsmanship: interviews (including Doug McIlroy), a 57‑byte ELF teardown, a 440‑byte metamorphic ELF-64 virus writeup, and pieces on polyglots, kernel loading, and packing. The aesthetic matters: tight, example-driven work that shows how big systems are often debugged into existence, and why tiny, deliberate artifacts are excellent teaching tools. Read the issue on tmp.0ut Volume 5.
"How can people write big systems in general? We do, but as you say, they're often debugged into existence rather than cleanly designed into existence."
RotaryCell: Rotary phone over LTE with ESP32‑S3
Why this matters now: RotaryCell shows makers how to modernize vintage telephony safely and reversibly using widely available hardware and clear BOMs.
RotaryCell converts an unmodified rotary desk phone into a battery‑powered cellular handset using an ESP32‑S3 + A7670 modem board, a telecom SLIC, and a protected 21700 cell. The project preserves the original dial, ringer and handset, supports pulse‑dialing, OTA updates, and exposes a Wi‑Fi maintenance interface — but the repo is explicit that this is an engineering archive, not a polished kit. The build is a great reference for anyone wanting to bridge old electromechanics and modern radios; the repo and wiring are on GitHub. Hacker comments focused on safety (don’t expect RJ‑11 to survive mains) and audio tradeoffs between old handset speakers and modern codecs.
EFF urges Newsom to veto AB 1709
Why this matters now: The EFF says California’s A.B. 1709 would force invasive age verification and cut off under‑16s from social platforms, creating immediate privacy and free‑speech tradeoffs.
A.B. 1709 would ban “addictive features” for under‑16 users, effectively removing recommendation mechanics and basic social signals from services for minors. The EFF warns the law could push platforms toward government‑grade ID checks or biometrics and disconnect vulnerable youth from support networks. The organization’s take, and the messy policy tradeoffs it highlights, are laid out in their letter to Governor Newsom at the EFF site. Expect litigation, implementation headaches, and heated debate about whether blunt bans are the right lever.
Deep Dive
44% on ARC-AGI-1 in 67 cents
Why this matters now: A solo researcher’s low-cost recipe for hitting 44% on ARC-AGI-1 challenges assumptions about model size and sample efficiency and will force rethinking of benchmark rules and transductive evaluation.
A single author reports training a very small autoregressive transformer and, with cheap test‑time training, achieving 44% on the ARC‑AGI‑1 benchmark for roughly $0.67 of compute and about 1.5 hours on an RTX 5090; full details and code are available in the original post. The gains aren’t from a bigger network — they come from a sequence of engineering and representational choices: SwiGlu activations, RMSNorm, flash attention, variable‑length training, stronger shuffling and augmentations, and crucially, 3D RoPE positional encodings plus per‑task embeddings. The author also switched the objective to train only on output tokens (an essentially supervised objective), noting the paradox:
"This approach is now supervised. What’s weird is that the test loss is now worse, yet it scores better!"
That oddity is central to the debate. The method relies on test‑time adaptation: the model sees test inputs (but not answers) and adapts at inference — a form of transductive evaluation. In one line: transductive setups let a model tailor internal representations to the specific test instance distribution, which can be powerful on metalearning tasks like ARC but raises the classic frame‑problem question: does success mean the model really learned general reasoning, or just how to exploit dataset structure?
The author’s ablations show that the positional and per‑task embeddings are the biggest levers — suggesting that representational tweaks can buy more than just scaling. Because the repo and experiments are open, others can reproduce, push back, or propose benchmark rule changes. What to watch: whether independent replication replicates the low cost and whether benchmark organizers change evaluation guidelines to limit—or at least label—transductive protocols. For ML researchers, the takeaway is clear: sample efficiency still has unrealized potential, and small, well‑tuned models plus smart inference tricks can move the needle dramatically.
Io_uring Without Readahead
Why this matters now: Turso’s measurements show that using O_DIRECT with io_uring can unintentionally kill throughput unless you reintroduce readahead at the application level — a practical lesson for anyone building storage backends.
A throwaway PR to Turso prompted a concise, measurement‑first writeup about how io_uring, O_DIRECT, and kernel readahead interact; the author’s post is at frn.sh. The setup: Turso supports two backends — a buffered syscall path using pread(2) and an io_uring path that opens files with O_DIRECT. O_DIRECT bypasses kernel page cache and, importantly, disables kernel readahead. Without readahead the io_uring path kept only one outstanding read at a time, so the block layer couldn't merge adjacent requests; adding a simple 32‑page application readahead window caused device requests to drop from ~196k to ~16k and the disk’s request merging (%rrqm) jumped into the 90s. The author sums it up bluntly:
"O_DIRECT takes away kernel readahead, so getting it back means implementing it in the application."
There’s a cost. O_DIRECT avoids kernel copies — which reduces CPU for copies but also skips the cache warming that buffered reads provide, producing more cache misses. Using sqpoll (a kernel thread that polls submission queues) further burned CPU on the test box; turning off sqpoll reduced system CPU substantially, but sqpoll helps only when you have spare cores. The post lays out the core tradeoffs: merged device requests and fewer syscalls versus higher CPU and cache miss behavior.
Practical takeaways: measure for your workload. If you care about device request merging and sequential throughput, application readahead with O_DIRECT can win. If you have lots of random reads or limited CPU, buffered pread or RWF_DONTCACHE may be a better fit. The post is a nice reminder that modern I/O stacks are not magic — small changes in where readahead happens can flip performance by an order of magnitude.
Closing Thought
Today’s standout pieces share a theme: clever, local engineering — whether a representational trick in a tiny transformer or a small readahead patch in a storage path — can unlock performance far beyond raw scale. Those wins force two honest follow‑ups: replicate the work, and decide whether your benchmarks and production decisions treat those tricks as legitimate progress or edge cases to be constrained. Both steps are healthy for engineering and for science.