flâneur

Kevin Liu

19 followers · 12 following · 1317 views

on the atlas — 38

highlights — 105

  • A computational system trained on electromagnetic measurements may have predicted these results perfectly, but would never have found radio.
    Designing AI for Disruptive Science - Asimov Press
  • detail only gives you more of the same kind of information — more roads, more mountains, more villages — when what you might need is a completely different schematic.
    Designing AI for Disruptive Science - Asimov Press
  • assuming we’re the exception who can override signals indefinitely. In practice the bill comes due somewhere, and it’s often not output that breaks first. It’s relationships, physical health, or mental health
    You're not burning out because you're tired — EA Forum
  • One underrated failure mode is treating burnout as a rest problem when the real issue may be perceived “trappedness” (identity, indispensability, heroic responsibility, community norms). That gets further locked in by a false A/B: either keep going as-is or quit. Naming that false dichotomy matters, because it blocks the moves that change the equilibrium. A question I’ve found useful: “What’s a third option?”
    You're not burning out because you're tired — EA Forum
  • Figure out what you're missing and ask for it. What do you need from your work to feel nourished? Autonomy? Positive feedback?
    You're not burning out because you're tired — EA Forum
  • There’s a general lack of cultural awareness in the Bay Area. It’s easy to hear at these parties that a person’s favorite nonfiction book is Seeing Like a State while their aspirationally favorite novel is Middlemarch.
    2025 letter | Dan Wang
  • I’ve learned that Christmas is a good time to write. Emails stop and all is calm. I submitted my manuscript this time last year in Vietnam.
    2025 letter | Dan Wang
  • I want to constantly try new things and expose myself to new information whether I enjoy it for not, because if I haven’t tried it before, how will I know for next time?
    You're Overthinking Packing - by Isabel Slone
  • In other words, you must be tranquil as a forest, but on fire within. In last year’s letter I went on about how you can only do one thing well at a time. Now I will tentatively venture out into two. I’ll admit there’s no way to adjudicate between values, and try to hold contradictory ideas in mind at the same time. Those ideas can be mundane. Work hard, but have slack for creativity. To be a good scientist you have to be both arrogant and humble. Care about everything, at the same time that nothing really matters.
    2025 letter | Zhengdong
  • Work very hard, certainly, but even Sergey Brin pinpoints 60 hours as the sweet spot (976? 403?), and Sam Altman’s first advice is facing the right direction, if you don’t want to take it from me
    2025 letter | Zhengdong
  • Many short trips also aligns perfectly with my advice to pack lighter, by the way. Just learn how to use flights well. It’s the perfect place to do deep work or reflect
    2025 letter | Zhengdong
  • As such, you can take a barbell strategy to travel. Either go for a year, or two, or ten, however long you need to go to open a bank account. Or, stay no longer than a few days in the same place, and come back often.
    2025 letter | Zhengdong
  • Specifically, you should allocate some fraction of your time towards improving your general knowledge of ML as opposed to working on your current project. If you don’t allocate this time, then your knowledge is likely to plateau after you learn the basics that you need for your day-to-day work. It’s easy to settle into a comfort zone of methods you understand well—you may need to expend active effort to expand this zone.
    An Opinionated Guide to ML Research
  • Every 1 or 2 weeks, I do a review, where I read all of my daily entries and I condense the information into a summary. Usually my review contains sections for experimental findings, insights (which might come from me, my colleagues, or things I read), code progress (what did I implement), and next steps / future work
    An Opinionated Guide to ML Research
  • During your day-to-day work, you’ll make incremental improvements in performance and in understanding. But these small steps should be moving you towards a larger goal that represents a non-incremental advance.
    An Opinionated Guide to ML Research
  • (albeit with a narrow tool in a narrow domain)
    Dario Amodei — Machines of Loving Grace
  • a surprisingly large fraction of the progress in biology has come from a truly tiny number of discoverie
    Dario Amodei — Machines of Loving Grace
  • It is by speeding up the whole research process that AI can truly accelerate biology
    Dario Amodei — Machines of Loving Grace
  • There is fiscally conservative reticence to expand government spending in the long run, especially in light of stories in the Boston Globe shaming workers who, through overtime, earn atypically large wages, leaving the impre ssion that those wages are common f or the public sector
    Final Boston Case SRF
  • the MBTA’s energy efficiency manager earns $85,000 and could make twice as much in the private sector (according to the official, the manager is only staying to put in the number of years required to earn a full pensio n when they retire
    Final Boston Case SRF
  • After months of s tudy, this is exactly what the consultant found.
    Final Boston Case SRF
  • an informed professional committed to keeping costs down should have immediately explained t hat , as GLX is a light rail extension operating in an existing right - of - way with active commuter rail , building a tunnel would be costly and redundant
    Final Boston Case SRF
  • B y 2018, there were 83 full - time employees working on GLX. During the first iteration of GLX, as a point of comparison, it was reported that on ly four to six full - time MBTA employees managed the multibillion - dollar project
    Final Boston Case SRF
  • B y 2018, there were 83 full - time employees working on GLX. During the first iteration of GLX, as a point of comparison, it was reported that on ly four to six full - time MBTA employees managed the multibillion - dollar project.
    Final Boston Case SRF
  • t, “ When [CM/GC] works well, it is us [the agency and all of the contractors] against the project” (Interview C 2020). Design - Bid - Build, by contrast, he described as extremely confrontational and riven with bitterness because each contractor tries to protect i ts liability and offload risk onto the agency or subcontractors
    Final Boston Case SRF
  • those working on GLX decided to apply the agreement to the ir new stations as well, designing stations with redundant elevators, escalators, full enclosures, and fare arrays rather than a platform with a partial weather shelter, as was initially pla nned
    Final Boston Case SRF
  • In an effort to mitigate the negative air quality impacts of the Big Dig, Governor Michael Dukakis committed to several transit projects, i ncluding completing GLX by 2011, in order to comply with the Clean Air Act
    Final Boston Case SRF
  • Los Angeles County’s preference for outsourcing planning to private consultants with little public oversight works well for simple projects like parking but not for more complex ones like urban rail.
    Final Boston Case SRF
  • American road projec ts are essentially commodities (Interview A 2020 ) . For example, a new public parking garage would be one of thousands of such structures built, which means that the costs and risks are well - known.
    Final Boston Case SRF
  • The more typical investment in roads is a bypass here, a new interchange there, and a widening yon der, all repeated hundreds of times to produce hundreds of billions of dollars in roadway expansion per six - year transportation bill cycl e.
    Final Boston Case SRF
  • put in other words, the economic gains from being able to build dense urban transportation networks are likely to be about 10% US - wide
    Final Boston Case SRF
  • Anderson had a dark gray balloon manufactured by Goodyear Tire and Rubber Company. A 25-foot (7.6 m) cube-shaped balloon was somewhat unusual, but the Manhattan Project's AAA priority rating ensured prompt delivery with no questions asked.[63][77]
    Chicago Pile-1 - Wikipedia
  • a nuclear reactor is likely to remain an especially complex and expensive capital project, and at our current current technology and safety requirements is unlikely to result in significantly reduced electricity costs.
    Why Does Nuclear Power Plant Construction Cost So Much? - Institute for Progress
  • The argument for nuclear ships in the Navy has never been “nuclear is a cheaper source of power” (though it theoretically could be if oil prices rose enough). Instead, proponents have argued the benefits of nuclear propulsion, namely unlimited operation at high speeds without worrying about refueling
    Why Does Nuclear Power Plant Construction Cost So Much? - Institute for Progress
  • Emphasis is thus placed on maintaining the industrial base that can construct them. For instance, the second and third Seawolf subs were built largely to maintain continuity in the industrial base until the next submarine program began.
    Why Does Nuclear Power Plant Construction Cost So Much? - Institute for Progress
  • During the construction of Vogtle's first two units, capital investment required jumped from an estimated $660 million to $8.87 billion.
    Vogtle Electric Generating Plant - Wikipedia
  • DEL’s, or DNA-encoded libraries, to study objects of interest. By leveraging the fact that the scientific community can cheaply sequence DNA at ridiculous scales (>trillions of nucleotides a day), and finding clever ways to tie their experiments to DNA, these companies can achieve previously impossible levels of scale in data collection
    Wet-lab innovations will lead the AI revolution in biology
  • After all, scale alone on those datasets haven’t yielded especially impressive results; independent replications of Alphafold2 find that it could’ve used 1% of its input dataset and still achieved near-identical accuracies.
    Wet-lab innovations will lead the AI revolution in biology
  • I imagine many saw this as the beginning of what happened in NLP. Politely, but firmly, showing domain experts the door. Get those linguists out of here, more data will replace whatever insights they have!
    Wet-lab innovations will lead the AI revolution in biology
  • Even benchmarks directly crafted from the problem are challenging to create. There could be huge motif overlaps between training/test cases (unnoticed unless you directly look for it)
    Six things to keep in mind while reading biology ML papers
  • because the ML community prizes novelty fairly high, this can create a feedback loop where poorly-motivated (but unique) approaches that don’t address real problems in biology consume the lion’s share of the ML attention economy.
    Six things to keep in mind while reading biology ML papers
  • es, which ends up yielding a far more generalizable model. Unfortunately, this also means it performs worse on the benchmark. Highly recommend reading their post!
    Six things to keep in mind while reading biology ML papers
  • Keep in mind that benchmarks in this field are incredibly challenging to create; the distribution shifts created when moving across datasets are massive.
    Six things to keep in mind while reading biology ML papers
  • telomere length doesn’t actually seem to be predictive of much, so clocks based on it are suspect
    Some questions (and answers) I had about longevity research
  • methylation status of 66 to 75 percent of methylation sites are driven by random processes
    Some questions (and answers) I had about longevity research
  • Clocks based on epigenetic ages aren’t dissimilar to this. Clearly, they do correlate with age-related cellular dysfunction. But, depending on how the clock is constructed, they can also be driven by entirely time-based stochastic processes unrelated to cellular dysfunction
    Some questions (and answers) I had about longevity research
  • The most commonly discussed biological clocks are epigenetic clocks. The chemical markers that cover your DNA — or epigenetics — affect how it’s converted to RNA, and thus, to proteins.
    Some questions (and answers) I had about longevity research
  • The main issue is just that the sample size wasn’t large enough: only 8 control mice and 12 drugged mice. Overall, it’s an excellent and well-supported result.
    Some questions (and answers) I had about longevity research
  • The difference in approach led to a complaint from the United States that Canada had pegged the world price of wheat.[1]
    William Craig McNamara - Wikipedia
  • Farmers got a guaranteed price for that crop, paid immediately, and later a further payment once the Board had completed the year's sales. This system of guaranteed prices and distributed income was extremely popular and when the Board dissolved in 1920, many farmers were livid.
    Canadian Wheat Board - Wikipedia