spelk

Not paying back US debt triggers a global depression.

In Commonwealth countries, it's possible to initiate a private prosecution (by a person who hasn't received criminal justice). These don't usually get very far, but I honestly wish these could be used to prosecute people who are treated differently for the same crime because of their status, or at least set a precedent that such activities are definitively okay.

Admittedly a controversial take, but I think some of the AI skepticism I see is driven by status anxiety, alongside the perfectly legitimate complaints about verbosity, low-effort output, lack of a human voice, etc.

The bar for what counts as expertise is shifting. If someone with relatively little background can use an LLM to get to a reasonably informed answer much faster than before, then experience and credentials alone carry a bit less signalling power than they used to. That obviously doesn't mean the inexperienced person suddenly has the judgement, context or intuition of an actual expert, but I do think some of the hostility toward AI comes from discomfort with that change.

AI is still wrong, goes in weird circular loops, and doesn't solve the fundamental human problems around judgement, organization, learning curves, accountability, etc. But it does seem to be raising the baseline expectation for what a competent knowledge worker should produce, and I feel that shift is genuinely uncomfortable for a lot of people who based their identities around being rigorous abstract problem solvers.

>3.7 Flash is available through the end of the year at an introductory price 1 of $0.75/1M input tokens and $3.75/1M output tokens. This price combined with the enhanced model performance enables developers and customers to scale production-ready agents cost effectively.

Introductory pricing until December 2026 implies no significant Gemini Flash developments until the next year.

My university let us take 3-4 courses, up to 12 credits total, as Pass/Fail, but only for electives.

I ended up taking just one that way. The policy was really designed for the breadth year because in upper years, you had much less room for electives, but you could try a difficult or unfamiliar subject without worrying too much about your GPA. If you discovered you liked it enough to switch into that major, the course would then be converted to a regular graded course.

I think allowing 3-4 Pass/Fail courses in the first year strikes a good balance. It gives students room to challenge themselves and explore unfamiliar subjects, while still preserving grades as a useful signal of whether they're prepared for college-level work.

Is the assumption here that inference costs will stay roughly static, or that frontier models will keep getting more expensive quickly enough to offset efficiency gains?

Because I don’t think “the subsidization train is going to end” necessarily means current pricing becomes impossible.

If capital keeps pouring into frontier AI, companies still have an incentive to subsidize access while competing for users and market share. And if that subsidization starts drying up, there’s even more incentive to bring inference costs down by making smaller and cheaper models catch up to today’s frontier capabilities.

So either way, I’m not sure you can extrapolate from the cost of serving current frontier models to what equivalent capability will cost two years from now.

I think like you mentioned, the practical reasons are disproportionately oriented around either privacy (, a clear constrained workload (need to OCR files/transcribe audio, and there isn't really a clear or meaningful reason to switch out the model to chase new incremental gains), or regulatory compliance (e.g. source code, patient data, can't leave the country).

I don't think this would make a big difference for the threat model described in the OP? They'd still be able to use your IP Address and potentially do nefarious things through your role as an unwitting proxy.

I think from a technical perspective, newsletters are a solved problems with lots of options as you've mentioned.

Substack thrives because it's built a reputation for having well-reasoned or engaging voices in a newsletter format. They pushed a big campaign to onboard writers like Bill Bishop (Sinocism) onto the platform early on. Because of that, newer writers are likely seeking to tap into the network effects/discovery systems that comes with having big names on the platform.

People log onto Instagram, Tik Tok, X, etc. because the people they're interested are already on the platform, and the platform provides a recommendation system that allows newer "creators" to tap into an existing audience for others in their niche.

I am quite sad that Kagi Smallweb excludes Substack. I understand that less technical writers have started moving to Substack both to build an audience and have access to effective publishing tools, but it feels like I'm missing out on some really interesting new perspectives, similar to early Medium.

This is fantastic work, really impressive is an understatement. I really hope this sets a new DeepSeek-esque standard and starts another the death knell for companies continuing to cosplay as frontier labs (like Cohere).

I miss Google Text ads. I supported AdBlock Plus's decision to exclude them from default filtering. Man, what happened?