seangoedecke-com

C2PA digitally signs images with cryptographic authenticity metadata, but is currently ineffective because almost no human-taken photos are signed while most AI images are—making signature removal a trivial way to fake human origin. For the system to work, all cameras must sign by default and all platforms must preserve metadata, requiring years of effort and regulation.

Blogging about unfamiliar topics forces deeper research, clearer thinking, and learning through writing. The author recommends taking a controversial stance in each post, changing your mind while writing, and using feedback (including from LLMs) to improve, arguing this practice is valuable even without publishing.

The EU AI Act requires AI text outputs to be watermarked by August 2026. Methods like Google's SynthID and Unicode homoglyphs exist, but all text watermarks can be trivially removed via paraphrasing or character replacement, making them fundamentally different from image watermarks.

Stating the obvious in writing is surprisingly useful—it clarifies thoughts below conscious awareness and offers catharsis for those who notice truths nobody discusses. Though hard and sometimes risky, articulating the obvious lets writers build toward nuance and helps newcomers feel validated.

The article argues AI inference is profitable, with estimated costs around $1 per million tokens and API prices far higher, supporting 70-80% gross margins. AI labs charge high margins to subsidize training, but inference itself remains a viable business even if those labs fail.

Pipelines use code-defined control flow for LLM programs; agents let the LLM manage control flow with tools. Pipelines are predictable but rigid, while agents are smarter and more flexible at the cost of unpredictable latency and cost. The author recommends using agents when in doubt.

A developer shares personal projects built with AI assistance, including a daily SkiFree puzzle game, an auto-generated Anki card tool, an endless wiki, and an offline plant identification app. He argues these niche projects would not have existed without LLM help, showing AI enables software that previously wasn't worth the effort.

The article argues that some anti-AI rhetoric among programmers mirrors fascist tropes: nostalgia for a purer past, rejection of modernism, and defense of an elite fraternity. It compares this to fascist writings and notes the Luddites shared similar conservative grievances. While not calling all anti-AI views fascist, it warns against treating AI as an absolute enemy.

Software engineers should deliberately work at 80% capacity to stay available for high-impact, time-sensitive tasks. Doing nothing creates space to notice opportunities, avoid burnout, and resist low-value glue work or exploitation. Strategic slack makes it easier to perform well on what truly matters.

Engineers and product managers often struggle to build trust due to non-overlapping skillsets and frequent broken commitments. To improve the relationship, engineers should understand the PM's perspective, be consistently accurate, defer on political calls, and accept that some PMs simply won't work well with them.

The article disputes the claim that AI inference GPUs last only three years, tracing it to an anonymous, incentivized source. Evidence like Google's eight-year-old TPUs, AWS never retiring A100s, and supercomputer data showing over 95% GPU survival at three years suggests GPUs can last much longer physically, even if economic lifespans are shorter.

A benchmark of 200 images found that OpenAI's elaborate "GeoGuessr" prompt did not improve o3's geolocation accuracy over a basic prompt—it performed slightly worse. The author warns against overestimating prompt engineering based on anecdotal success, and notes o3's geolocation skill has not carried over to newer GPT models.

Prompts for AI coding tools are a form of technical debt that decays silently with each model upgrade, unlike code. The author advises against investing heavily in bespoke agentic setups, recommending instead using third-party tools with minimal configuration and keeping custom prompts limited to concrete project facts.

The "just-say-no engineer" who blocks changes to maintain quality was a product of the ZIRP era, when tech companies had bloated teams and little pressure to deliver. With ZIRP over, companies now prioritize shipping features, making this role less valued. The article argues this shift is often blamed on AI but would have happened regardless.

An engineer describes the shift from 2025 to 2026, where AI agents are now reliable enough to write entire PRs, diagnose most bugs, and handle testing and setup tasks. He avoids using LLMs for public communication like PR descriptions or Slack messages, and emphasizes finding the right balance between over- and under-utilizing agents.

DeepSeek-V4-Flash and DwarfStar 4 make LLM steering—directly manipulating model activations mid-inference—practical for local use. The author is skeptical about its utility, arguing most gains can be replicated with prompting or fine-tuning, but expects the next six months to reveal if steering has real applications.

The article examines Elon Musk's proposal to build AI datacenters in space. While cooling via heat radiation is feasible—requiring roughly 250,000 square meters of radiators for a 100MW facility—it remains wildly impractical overall due to the massive launch mass needed, repair difficulties, and other logistical challenges. Cooling itself is not the fundamental obstacle.

Thinking Machines released Interaction Models, a fully-duplex voice system using 200ms micro-turns for more natural conversation, with a separate reasoning model bolted on for complex tasks. Its main achievement is scaling a fully-duplex model to include video input and far more parameters than previous systems like Moshi.

The article outlines progressive arguments for AI: LLMs aid disabled and chronically ill people, level class-based communication barriers, provide equal access to tutoring, and could advance a left-wing technological utopia.

AI coding tools like Claude Code have raised the floor for weak software engineers, turning net-negative contributors into merely mediocre ones. Some engineers now act as thin wrappers around LLMs, adding little value themselves, which may threaten their long-term job security as companies assess what value engineers add beyond AI.

Most incidents are boring and resolve on their own; impulsive fixes often worsen them. The recommended first step is to do nothing, and effective actions are usually simple, like disabling a feature flag. While resolving incidents earns political credit, it does not build lasting power since executives cannot assess the effort involved.

AI progress hasn't slowed despite longer training horizons, possibly due to huge FLOP efficiency gains from fixing bugs, unreliable human intuitions about near-human intelligence, and capabilities depending on traits beyond intelligence like persistence.

A staff engineer argues Will Larson's famous "staff engineer archetypes" are descriptively useful but bad goals. Instead of aiming for an archetype, engineers should focus on building trust and being useful to the company, delivering value regardless of circumstances.

Even if AI makes software engineers worse over time by reducing learning, they may be forced to use it for short-term gains or be outcompeted. The author compares this to pro athletes with short careers, suggesting software engineering may no longer be a lifetime career and engineers should plan accordingly.

Notes comparing historical Luddism to modern anti-AI movements. Luddites were skilled weavers with concrete local goals who enjoyed near-universal class solidarity but were ultimately crushed by the state. The author argues Luddism is a poor model because AI models are intangible information, not smashable machines.

The article examines historical Luddism as a decentralized 19th-century movement where skilled textile workers destroyed machines automating their jobs. It draws parallels to modern anti-AI activism but notes key differences: Luddism was local and targeted specific machines, while AI concerns are global and datacenters can affect jobs worldwide.

The article argues that only three types of AI products currently work effectively: chatbots like ChatGPT, completion-based products like GitHub Copilot, and agentic products like coding agents. It suggests AI-generated feeds and AI-based video games are promising but not yet successful product categories.

Evaluating new AI models takes months because standard benchmarks are unreliable and often gamed by companies. Real-world testing requires significant time and effort, while subjective "vibe checks" provide limited insight. This makes it difficult to determine if AI progress is stagnating or if models are genuinely improving.

The article provides strategies for software engineers to avoid work blockers, including working on multiple tasks, sequencing work to minimize blockers, maintaining reliable developer tooling, debugging outside one's area, building relationships with other teams, and leveraging senior managers for support.

Big tech companies produce sloppy code because engineers often work outside their expertise due to short tenures and frequent team changes. Most code changes are made by relative beginners unfamiliar with codebases, while experienced engineers are overloaded. Companies prioritize flexibility over long-term expertise, accepting bad code as a tradeoff.