Today’s Highlights#

The best place to start is Google’s Gemini 3.7 Flash announcement. Pitched as a workhorse for coding and agents, it promises steadier multi-step execution and a lower price for running those loops at scale. That idea — make capability usable, not just stronger — ran through the day. DeepSeek Harness reframed the whole agent as a set of swappable plugins you can trace and replay; Cerebras and OpenAI previewed a wafer-scale path to near-real-time frontier inference; and a different kind of infrastructure made its claim too, from an open financial terminal to a new yuan clearing rail in Frankfurt and a 50 TB public archive locked inside someone else’s data center. Speed ahead, rules and memory close behind.

Tech and Products#

Stronger models matter when they fail less, cost less to run, and leave a trail you can follow.

Gemini 3.7 Flash: a steadier workhorse for agents#

Google says Gemini 3.7 Flash is its most capable Flash yet for coding and agents, arriving three weeks after 3.6 Flash. Compared with its predecessor, the company reports higher first-pass accuracy and better production-ready generation on FrontierCode and DeepSWE, plus gains in web and UI generation (WebDev Arena Elo 1588 versus 1538) and in knowledge-heavy work across finance, law, and biosciences. The stated change is not only benchmark results: the model is meant to handle blockers, clarify intent, and follow instructions more faithfully, with more disciplined multi-step planning and tool use. It will power Gemini Spark for Google AI Pro and Ultra subscribers, improving Workspace tool use, and is offered through year-end at half the prior introductory price.

Note: Hands-on HN notes focused on vision-to-code tests. Readers found Gemini still strong at turning images into code for the price, though whether it consistently beats the top rival remained debated. The other thread was about pricing that is set to double after the introductory window. Many commenters treated it as a subsidy and warned that teams rarely switch models once integrated, so migration friction matters as much as the sticker price.

Discussion: Hacker News thread

DeepSeek Harness: everything is a plugin, every run is traceable#

DeepSeek’s Harness developer preview leans into a single idea: every agent capability is a plugin. Built on the Cordis kernel, models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI are all plugins that can be swapped or recomposed in configuration without forking the harness. The companion promise is traceability. Everything the model sees — system prompts, reasoning, tool calls and results, scheduling, context injections — is appended to a session log that can be inspected by source, resumed, forked, searched, and replayed. Four presets sketch the range: Standard with the full toolkit, Code mode that lets the model orchestrate multi-step work in TypeScript, a Minimal two-tool rig for benchmarking, and a Creator mode for inspecting the runtime and trying new plugins.

Note: HN spent less time on feature lists than on tradeoffs. A long thread asked why so many harnesses pick Node.js; advocates cited async and distribution, critics warned about supply-chain and footprint. Another strain pushed back on instructing a model to do what deterministic code already does well — for example, running pre-commit checks via a skill instead of a git hook — and asked when a prompt should stay a program.

Discussion: Hacker News thread

Spaghettifying DRAM: when the last address map becomes a switch#

Security researcher Christopher Domas released skitter-creek-bath-salts, a project that rewires the final stage of the memory pipeline. By flipping configuration bits in the DRAM controller, it temporarily scrambles the translation from physical addresses to DRAM coordinates, so protections built above that layer no longer line up with where bits actually land. Demonstrated on AMD Family 16h, where the relevant registers are documented, the trick uses a single XOR on a controller register alongside careful cache and timing control to enter a scrambled view, fetch from a protected region, and restore normal mapping. Because the exact transform is underspecified, the work collects alias pairs by planting sentinel values in one view and scanning for them in the other, then solves the linear map and derives aliases for any protected address.

Note: Commenters looked forward to the promised Black Hat talk while debating the project’s dramatic README. Some readers found the voice — which several attributed to AI assistance — overblown and tiring; others thought the pipeline diagrams earned their emphasis. On substance, the thread stressed that the risk is structural: channel, rank, and bank interleaving exists on many modern controllers, so later generations simply hide the knobs.

Discussion: Hacker News thread

Business and Platforms#

Speed and cost set what products feel like; open tooling decides who gets to build them.

GPT-5.6 Sol Ultrafast: frontier speed without a quality cut#

Cerebras and OpenAI previewed Ultrafast, a tier that runs GPT-5.6 Sol on Cerebras wafer-scale hardware. The headline claim is up to 750 output tokens per second with no quality drop, about 11 times faster than Fable 5 and five times faster than Opus 4.8 on Fast mode. As a rough yardstick, Cerebras reports answering 2,500 Humanity’s Last Exam questions in about 11 hours versus more than three days for the comparator. The hardware story is 44 GB of on-wafer SRAM so weights stay on chip and tokens stream through pipelined layers, avoiding repeated off-chip transfers. OpenAI frames the win as putting agents on the critical path for outages and security response, with live preview for a small set of customers and wider access as capacity grows.

Note: HN focused on what “fast” should mean. Some readers argued that 2,500 independent questions could be parallelized, so single-answer latency matters more. Others said speed changes attention: less waiting, fewer context switches, and the chance to use agents inline rather than in a queue. Practitioners added a cost angle, noting Sol’s token efficiency in their workloads, while remaining cautious about Cerebras pricing and capacity.

Discussion: Hacker News thread

Gloomberb: an open, keyboard-driven finance terminal#

Gloomberb is an open-source terminal for markets, available as a desktop app and a text interface. It is command-bar first: type a ticker or a shortcut such as DES for security details or TOP for ranked stories and jump straight to the view. Panels cover quotes and charts, financials and filings, holders and insiders, options and analyst ratings, global indices and FX, yield curves, movers, and sentiment, plus workspaces for portfolios, watchlists, broker links, alerts, notes, and chat. The site shows about 1.3k GitHub stars and highlights install via a single shell command alongside direct binaries, with a look that deliberately nods to the Bloomberg terminal while aiming at individuals and small teams.

Note: HN was clear that a terminal is more than its chrome. Bloomberg’s moat, commenters argued, is data, news, and the chat network that settles most bond trading; a $32,000 annual seat can pay for itself on one good trade. Readers therefore placed Gloomberb closer to TradingView for retail use, or as a light aggregator, and debated the install script and JavaScript bundle size as practical adoption questions.

Discussion: Hacker News thread

Policy and Governance#

New rails need clear ownership, and archives need a second copy.

Frankfurt joins the yuan clearing network#

China has authorized Deutsche Bank to clear renminbi in Frankfurt, the first time a foreign bank has held that role in Europe, Traders Union reports. Until now European clearing ran through Chinese banks in Frankfurt, London, Paris, Budapest, and Luxembourg. The designation follows a memorandum with the People’s Bank of China and, according to the report, builds on a 2019 model that let Deutsche Bank route onshore renminbi FX through Hong Kong. For European firms in manufacturing, autos, and green technology with deep China links, the practical effect is another path to invoice and settle directly in yuan without first converting through a third currency. The piece places the move alongside wider steps to add direct participants to China’s Cross-Border Interbank Payment System, describing it as plumbing expansion rather than an immediate challenge to the dollar or euro.

Note: The thread’s consensus was modest. More pipes do not by themselves shift currency hierarchy; they lower friction for firms that actually trade with China and reduce single-path dependence. Other readers watched for second-order effects on European infrastructure and whether cost and convenience will pull more settlement into direct yuan.

Discussion: Hacker News thread

A public archive locked behind someone else’s contract#

St. Louis public broadcaster Nine PBS has sued Iron Mountain in Denver District Court to recover more than 50 TB of archival material — seven decades of local programming — held in an Iron Mountain data center. The station says it stored the archive through a vendor, Open Source Storage, whose site went defunct and whose corporate status lapsed. When the agreement expired on March 6, access was cut despite a 30-day retrieval window in the contract. The physical media, the complaint says, sits in Iron Mountain’s facility under equipment owned by OSS, and Iron Mountain declined to release it without direction from its customer. The station obtained a default judgment against OSS in St. Louis confirming ownership and a right to possess, but must still compel the facility operator in Colorado; a judge has granted temporary preservation.

Note: HN kept returning to basics: three copies, two media, one off-site. Many readers said digitization in 2019 should have triggered that discipline regardless of the archive’s age, while others noted nonprofit budgets and staffing constraints that push work to a single vendor. The thornier point, several commenters wrote, is custody: when the contract holder and the facility holder split, a data center cannot safely hand data to a third party without a court order.

Discussion: Hacker News thread

Science and Research#

Old definitions meet today’s tools: what a type meant in 1910, and what it means now.

Principia Mathematica, read as a programming language preface#

A close reading of Chapter 1 of Whitehead and Russell’s 1910 Principia Mathematica finds ideas that feel contemporary to programming languages: the distinction between intension and extension and what is now called referential transparency, with the non-transparent case “A believes p” as the classic counterexample; definitions as typographic convenience that nonetheless carry intent; propositional functions written with a hatted variable that anticipate lambda terms, with careful attention to free and bound variables, substitution, alpha-renaming, and scope; and early uses of the word “type” in its modern sense. The notes also draw out the any-versus-all nuance around schematic assertions and an intuitionistic flavor of existence, plus the move to treat functions as special binary relations. The author’s point is that if published now, the thousand-page proofs would belong in an appendix checked by a prover, while the core contribution would be the setup of basic notions.

Note: Commenters corrected the loose claim that Gödel found a “logical error” and instead emphasized limitations of sufficiently strong formal systems. Others linked forward to homotopy type theory and constructivist lineages, and spent time on the material history of “penalty copy” mathematics and how proofs move from ink on paper to proof assistants.

Discussion: Hacker News thread

Society and Culture#

When tools get faster, habits need choosing.

Choose boring technology — and why the phrase still stings#

Dan McKinley’s 2015 essay Choose Boring Technology returned to the front page and reopened an old argument. The image is innovation tokens: a company gets about three for a long stretch. Spend one on Node.js here, one on a young database there, and there is little left for the work that actually differentiates the business. The essay distinguishes known unknowns from unknown unknowns, and argues that boring, well-understood stacks — MySQL, Postgres, Python, Memcached — are valuable not because they are exciting but because their failure modes are. Etsy anecdotes show the payoff: building activity feeds on PHP and Memcached was harder up front, then scaled twentyfold while neglected because the shared platform stayed known. Add new tech only through a visible conversation, try without it first, write down what actually hurts, and commit to migrating old uses.

Note: HN split precisely where the essay predicted. Critics said “boring” is a weak proxy and asked for crisper tests like tested, operable, and familiar to the team. Supporters said the metaphor mattered in a period of frothy adoption and tamed resume-driven choices. Several added that boring is relative: what was edgy in 2015 can be boring now, which is the point — boring is about what your team already understands, including how it breaks.

Discussion: Hacker News thread

Ordinary abundance: a walk through the apartment we take for granted#

Ordinary Abundance turns a single evening at home into a museum. Scrolling through an apartment, the piece pairs each comfort with the awe it once caused: music on demand imagined as the limit of felicity, electric light that turned a courthouse square into midday, art made portable, portraits fixed in silver, spectacles that returned reading, and a telegraph that made waiting weeks for a reply seem quaint. In the kitchen, clean tap water draws cheers, a pineapple is priced like a sheep, ice travels from New England to Calcutta, and spices read as if from paradise. Later the night lists vaccines that erased smallpox, anesthesia that ended surgical agony, indoor plumbing, bicycles, washing days, sewing machines, and radiators — all presented through period voices who could not take them for granted.

Note: The thread circled hedonic adaptation. Some readers tried to practice noticing a hot shower or quiet room more deliberately; others proposed a “chaos monkey” day without one convenience. Several argued that abundance has not automatically lifted shared happiness, and a few bristled at the page’s parallax scrolling that, on narrow windows or in reader mode, covered the text.

Discussion: Hacker News thread

Closing#

From address maps to policy rails, the day was about plumbing. Making agents useful meant making them traceable and affordable; making frontiers useful meant making them fast enough to keep up with a person; making memory useful meant knowing who owns the box it sits in. An old logic book reminded us that careful definitions outlive clever proofs, and a single apartment reminded us that our ordinary already contains centuries of asks answered. See you next time.