rahimnathwani

Do they work well as a replacement for a monitor when you're travelling?

I don't know what records I'm missing. I usually only know of the existence of a record once I've received it.

"The work falls outside normal scope of the person's job."

Maybe in some cases, but not in this case with this agency (SFUSD).

The responses I've received so far from SFUSD (over several years) were each from someone whose whole job is/was dealing with public records requests.

One of the requests that was refused was for a Google Slides deck that was visible to all of SFUSD's ~10000 employees.

Some agencies have a different tactic: go really slowly.

I sent a public records request to my local school district on March 9th 2026.

They responded on 19th March 2026 saying they'd respond on in late April.

Since then, every month they've sent an email with some small portion of the records, with a PDF letter saying they'll send a follow up a month later.

I've received 8 of those responses so far, and there's no indication when, if ever, I will have received all the records responsive to my request.

In the UK, 'pavement' doesn't refer to any paved road, but to what Americans call a 'sidewalk'.

Bicycles are ridden on the road, not the pavement.

Sometimes people ride bicycles on the pavement, but doing so with a long piece of wood would be very dangerous indeed.

Yes, the reward needs to include both correctness AND the Brier loss. Sorry if that wasn't clear from my comment. I was focused on the bit OP said was missing.

This is marked on the page as an 'Original Investigation', but right below the title it says it's a 'Systematic Review and Meta-Analysis'.

I'm skeptical about meta analyses in general because, the times when I've dug into a meta analysis and traced the conclusions back to the original papers, I've always been disappointed, and ended up not being convinced of the conclusions of the meta analysis.

Anyway, I haven't done that exercise for this paper, but I will say:

- Although the authors say they did a "systematic review and meta-analysis of 19 studies", only 12 of the 19 papers are included in the meta analysis.

- Only one of the authors read all the studies they included ("Study quality was assessed by 1 reviewer.")

- That reviewer was Lucy Corrigan, BSc ("Full texts were screened (L.C.) against the same inclusion and exclusion criteria (165 articles)...")

Here is the abstract of one of the 12 studies they included (and marked as 'Good' quality, the highest rating they gave): https://www.tandfonline.com/doi/full/10.1080/19485565.2021.2...

Here is what they found:

  Discrete-time hazard analysis indicates that the log odds of mortality were significantly lower among 18 and 19 year old adults ever diagnosed with ADHD and significantly higher among 46 to 64 year old adults ever diagnosed with ADHD, with a crossover occurring at age 33.

If you found this interesting, you should read section 5 of this article to learn about how LLMs can be trained so that they're good at confidence scoring: https://magazine.sebastianraschka.com/i/217935364/51-on-cali...

OP says that LLMs were never scored against correctness. But that's an assumption about the optimization process.

This is sort of right:

- pretraining scores on next token, and says nothing about confidence

- RLHF and RLVR scores are binary (1=preferred/correct, 0=worse/incorrect) and the feedback doesn't do anything to push the weights in a direction that improves calibration/confidence

BUT you can ask the model to output its confidence, and then train it to be more correct. RLCR (mentioned in the article I linked above) adds a Brier penalty to the reward, so the model's performance isn't just measured on % correctness. (Presumably TypeSafeAI's RLCD is similar to RLCR, but they haven't published anything about it.)

This eliminates the "always say 30%" loophole OP mentioned. The Brier penalty, (q-c)^2, punishes a forecaster who ignores instance-level information. For example if we have a model that's right 60% of the time, it could use one of these strategies to give its confidence:

  Strategy A, always say 0.6:
  When right (60% of the time): (0.6-1)^2 = 0.16
  When wrong (40% of the time): (0.6-0)^2 = 0.36
  Expected penalty: 0.6×0.16 + 0.4×0.36 = 0.24

  Strategy B: the model can tell its answers apart. Half are 'easy' (right 90% of the time) and half are 'hard' (right 30% of the time), which still averages to 60%. It says 0.9 on easy ones and 0.3 on hard ones:
  Easy: 0.9×(0.1)^2 + 0.1×(0.9)^2 = 0.09
  Hard: 0.3×(0.7)^2 + 0.7×(0.3)^2 = 0.21
  Expected penalty: (0.09+0.21)/2 = 0.15

Both strategies are perfectly calibrated (60% is both the average confidence and thje average % of time it's right), but B gets a lower penalty. Of course these are not actual strategies used by LLMs. But I'm just trying to say why training can push the model toward optimizing instance-specific confidence.

"There were no end entity certificates issued out of the affected CA’s."

This seems like a weird statement given that anyone can issue certificates out of the affected CAs.

Imagine I have a link of this format:

https://cf.metaaiusercontent.com/q3-revenue-review-7b1e42f0

No one can guess that URL. The q3-revenue-review-7b1e42f0 part is only ever shared with the Meta's server (protected by TLS).

Now imagine the URL is instead:

https://q3-revenue-review-7b1e42f0.cf.metaaiusercontent.com/

The first step for my computer to open that link is to do a DNS query, which immediately shares the entire URL (minus the `https://` and the `/`) with my ISP's DNS server (and anyone sniffing traffic along the path, as DNS queries are usually sent in plain text).

He also talked with Dwarkesh Patel: https://www.dwarkesh.com/p/dominic-cummings

capable, tenacious, kind

There may be no self-driving wheelchairs today, but they will probably exist in a year or two. If/when they do exist, and a self-driving wheelchair were returning to its owner's home after being repaired or tuned up, would it be OK for a pedestrian to block its path, since it's not human nor currently carrying a human?

If an empty self-driving wheelchair were on its way to pick up its owner, would it be OK for a pedestrian to block its way?

Does humanity design CPUs 'such that human life is super valuable'?

Jev is trained towards "Calibrated decisions: answers with epistemically honest probabilities on System One tasks." [emphasis mine]

To me it seems like:

- Jev should be good at telling us 'what is'

- Jev isn't trying to tell us 'what should be'

e.g. if you want your software to assume each human life has the same value, then you should rely on Jev to tell you how many human lives would be lost in each scenario, and then your software can pick. But Jev is not trained on your value system, which is neither objective truth nor something that you can derive from the inputs.

I assumed you were talking about instance-based billing (what I've always used), but I think you're talking about request-based billing?

If you scale Cloud Run to zero and then send a request, won't you immediately wake it up and be charged until it powers down in at least 15 minutes?

Not available to me on gemini.google.com (Pro, USA).

It was your post yesterday that made me aware of granite-speech-5.0-470m-turboctc (which I guess you used as a comparator because it's already very fast).

Thank you!

  uv tool install mlx-audio
  mlx_audio.stt.generate \
  --model ibm-granite/granite-speech-5.0-470m-turboctc \
  --audio.wav \
  --output-path transcript

(Until now I've been using mlx-community/parakeet-tdt-0.6b-v3)

When I read it was multi-user, I got excited, but then realized that it assumes each user already has an iPhone and/or iPad.

It's a shame that iPhoneOS, iPadOS and Omarchy are designed to be single-user devices.

Android phones (at least Pixels) support multiple users, each with their own apps and data.

Right but the later questions are harder than the earlier ones.

From page 4:

  Problem difficulty increased across the experiment from one-step1 (problems 1–4) to two-step2 (problems 5–8) to three-step3 (problems 9–12).

From page 6:

  Problems increased in difficulty from one-step (problems 4-6) to two-step (problems 7-10) to three-step (problems 11-14) problems.

The fact that the AI group did worse on the harder problems than the non-AI group did on the easier ones doesn't tell us anything.

He compared techniques that used:

- BM25 only

- Jev only

- LLM + Jev

The latter worked best.

But I wonder whether LLM on its own might have gotten close. If you're going to use an LLM for each call then it seems reasonable to test whether or not adding Jev improves performance.

EDIT: I should have had more coffee before writing this comment. As make3 points out below, I forgot that in the LLM+Jev setup, the LLM only needs to run once per query, whereas in the LLM-only setup the LLM would run once per item/segment being evaluated.

The article title is not supported by the evidence presented in the paper (https://arxiv.org/pdf/2604.04721).

The evidence for reduced persistence is kinda weak.

The strongest experimental result in the paper is: if people use AI to help solve a series of problems, and then the AI is suddenly taken away, they perform worse on subsequent problems than people who hadn't been using AI for the earlier questions.

But:

- They're looking at really really short-term effects, like seconds or minutes. That's not really interesting if we care about long term effects (which is what the abstract says is important).

- What they found is totally expected and isn't something I'd worry about: it's kind of expected that if one guy warms up with 10 questions done himself, and the other guy warms up with 10 questions delegated to AI, then the second guy is going to do worse on the final 10 questions (because he didn't warm up).

Like many others, I was surprised by his 'cancer cure' take.

But if you ignore the three slides about medicine, the rest of it is interesting, reasonable and informative.

I'm curious whether you've checked out https://github.com/yazinsai/OpenOats and how Biscotti compares.

This reminds me of the Casio TV-21 portable/pocket TV, which likewise used ambient light: https://www.flickr.com/photos/42444189@N04/12369631603