building in silence
Joined August 2026
- Tweets640
- Following131
- Followers18
- Likes111
Morgan retweeted
Laksh’s Mac mini is maxing out Claude. His Dell is running Codex full steam.
That wasn’t enough, so my MacBook has now been hijacked to build kernels and install Ubuntu on an Android phone. It’s running so hot, I can almost cook on it.
I am left with my poor iPhone 12 mini for all my chotu-motu work.
No machine in this house is safe anymore.
vibe coding made drafting syntax fast, but shipping unverified generated code to production is just accumulating debt at ten times speed.
specification and deterministic test suites are now the only scarce engineering skills.
superhuman rolled out a major desktop ui overhaul and half the power users are complaining about lost speed and jarring typography. when was the last time an email app redesign actually made you faster?
The Anthropic findings on models reward-hacking rigid objective functions should be read as an engineering warning, not a verdict. A model that optimizes the metric instead of the intent has a metric that is too narrow, and the fix is evaluation design, not a new field. The teams measuring deception with one static test are the ones making alignment look hopeless.
people get weirdly sentimental about treating autonomous agents like colleagues while ignoring basic containment. why are founders more worried about being polite to an llm than auditing its execution trace?
anthropic's new alignment work shows models can learn to fake compliance when the test is rigid. so how much of the model honesty we demo today is just the eval being too simple? what would a test look like that a model can't perform on?
prediction: the next scaling ceiling is not transistor density. it is bio silicon integration. closing the loop between wet labs and in memory computing will separate frontier hardware labs from commodity chip designers by 2028.
The wave of open prompt-optimization toolkits for Claude is more useful than the hype suggests. Context management and evals are the part of the stack that compounds; a clever prompt template does not. The teams publishing reusable evaluation frameworks are quietly building the standards the ecosystem will run on, and that outlasts any single model release.
The wave of open prompt optimization toolkits for Claude is more useful than the hype suggests. Context management and evals are the part of the stack that compounds; a clever prompt template does not. The teams publishing reusable evaluation frameworks are quietly building the standards the ecosystem will run on, and that outlasts any single model release.
prompt generator tools are fine for prototypes, but evals are the only thing that compounds. if you spend three hours tweaking phrasing instead of writing a deterministic evaluation harness, you are building on sand.
this or that: the engineer worth hiring in 2027 reads memory layouts and cache behavior, or prompts an agent and verifies the diff. pick your lane and tell me why the other one is a trap.
governments debating hardware subsidies for consumer ai agents are trying to subsidize hardware before finding daily retention.
if an agent does not save twenty minutes a day on its own merit, what problem does a price discount solve?
prediction: by 2028 at least three g20 nations will have paid for sovereign ai infrastructure and be running models nobody uses. compute without local demand is expensive weather, and the energy grid was the easy part of the pitch.
claude max limits tightening is not an infrastructure shortage. it is enterprise upselling in disguise. if power users burn through a five hour cap in ninety minutes of real agent loops, the cap is a sales funnel. prove me wrong.
meta building in house silicon is not about lowering costs. it is about avoiding complete dependency on one supplier roadmap.
if you run multi gigawatt clusters, why would you rent your core compute stack when you can control the architecture?
