dmayle

3 дня назадdiscuss

You're right... I read the title too quickly... I'll have to look at Senko vs Softformer later...

3 дня назадdiscuss

Fun... This is something I actually care about...

I used to keep a version of whisperx around, because I think it's important to have not just transcription, but also timing and speaker identification (e.g. for subtitles)... It depends on pyannote, though, which has some wierd licensing (and is tougher to script the installs because of it), so I wanted to look at something that both had better transcription, and supported diarization (the speaker and timing). I decided on parakeet for the transcription with softformer (the diarization), but most of the available engines for it don't include softformer.

I coded up an OpenAI compatible server for parakeet-rs ( https://github.com/altunenes/parakeet-rs ) (which does support softformer) and I've been using it with OpenWhispr (a desktop app for transcription that handles all sorts of neat thing).

I'm doing CPU-only transcription (because I use my GPUs for other stuff and haven't gotten around to adding in the GPU-path), but it's incredibly empowering to be able to have local transcriptions at will.

dmayle

12 дней назадdiscuss

For that price you can put together a PC with 128GB of ram ($2000) and an RTX 5090 ($3600) and get 70-100 tokens per second instead of 45

dmayle

12 дней назадdiscuss

For me, the simple shift of moving to eating 0.7g of protein per pound of body weight every day did it... It was not enough on it's own (I also eat more fiber, less fat, less carbs, though less does not mean none), but that alone was a night and day difference, and enough to change the story.

dmayle

13 дней назадdiscuss

The issue isn't food addiction, it's a broken hunger response (broken by Howard Moskowitz intentionally in the name of profits)...

You have to change your food regimen completely (higher fiber, more protein, less sugar, less carbs, less fat), and that's tough to do when you're surrounded by options that aren't...

I think the real problem is that the symptom we're trying to treat is "overweight", and it's actually a two-stage problem... Fix the hunger response... and only then work on fixing the weight... Fixinig the latter without fixing the former means you'll always gain the weight back, fixing the former doesn't guarantee you lose weight (and is only temporary, if you're using drugs for it)... You have to go after both problems.

dmayle

13 дней назадdiscuss

I had the same experience, but not with GLP-1 drugs, but by upping my protein intake to about 0.7g per pound of body weight.

Night and day, stopped always being hungry... I've tried Noom before (eating highly filling, low calorie foods, but filling, not satiating), but that only worked while I was tracking (and always forcing myself to keep it up)...

Losing weight required work on top of that, but the protein just made my hunger response start working properly again.

dmayle

в прошлом месяцеdiscuss

That's because of a fundamental misunderstanding of what an LLM is. The only correct answer to "Why did you do it like this?" is that the specific combination of input text and RNG state caused this particular output. There's no reasoning to be had.

* EDIT * What's with the downvoting? That's a correct description of what happened. You can't ask an LLM why it did something and expect a coherent response, because there's no thinking chain, and no stored thinking state... At best, you can get a reconstruction of how the context relates to the output (basically a summarization of the context).

dmayle

в прошлом месяцеdiscuss

You can swap out your default lenses with multifocal lenses... I use multifocal contact lenses, and my wife and my mother both had the surgery. My wife got the panoptix (no need for glasses at all) and my mother got vivity (just need reading glasses). At night, there are halos with the panoptix lenses (same with the multifocal contact lenses), the severity is not always the same per-person, and it bothers some people more than others (interferes with night driving), but it's an option. Yes, it's another surgery, but depending on your ability to afford it, and the amount it bothers you, it is still an option. From my point of view (admittedly, with contact lenses), going from three different pairs of glasses (vision, vision+reading, plain contact lenses+reading) to contact lenses with no glasses at all was just unquestionably worth it.

dmayle

4 месяца назадdiscuss

I run two 1.5TB Optanes in raid-0 with XFS (I picked them up for $300 each on sale about two years ago). These are limited to PCIE 3.0 x4 (about 4GB/s max each). I also have a 64GB optane drive I use as my boot drive.

It's hard to tell you, because it's subjective, I don't swap back and forth between an SSD and the optane drives. I have my old system, which has a 2TB Samsung 980 Pro NVME drive (PCIE 4.0 x4, or 8GB/s max) as root, and a Sabrent rocket 4 plus 4TB drive secondary (also PCIE 4.0), so I ran sysbench on both systems, so I could share the differences. (Old system 5950X, new system 9950X3D).

It feels snappier, especially when doing compilations...

Sequential reads: I started with a 150GB fileset, but it was being served by the kernel cache on my newer system (256GB RAM vs 128GB on the old), so I switched to use 300GB of data, and the optanes gave me 5000 MiB/s for sequential read as opposed to 2800 MiB/s for the 980 Pro, and 4340 MiB/s for the Rocket 4 Plus.

Random writes alone (no read workload) The optane system gets 2184 MiB/s, the 980 Pro gets 32 MiB/s, and the Rocket 4 Plus gets 53 MiB/s.

Mixed workload (random read/write) The optanes get 725/483 as opposed to 9/6 for the 980 Pro, and 42/28 for the Rocket 4 Plus.

2x1.5TB Optane Raid0: Prep time: `sysbench fileio --file-total-size=150G prepare` 161061273600 bytes written in 50.41 seconds (3047.27 MiB/sec).

    Benchmark:
    `sysbench fileio --file-total-size=150G --file-test-mode=rndrw --max-time=60 --max-requests=0 run`
    WARNING: --max-time is deprecated, use --time instead
    sysbench 1.0.20 (using system LuaJIT 2.1.1741730670)

    Running the test with following options:
    Number of threads: 1
    Initializing random number generator from current time

    Extra file open flags: (none)
    128 files, 1.1719GiB each
    150GiB total file size
    Block size 16KiB
    Number of IO requests: 0
    Read/Write ratio for combined random IO test: 1.50
    Periodic FSYNC enabled, calling fsync() each 100 requests.
    Calling fsync() at the end of test, Enabled.
    Using synchronous I/O mode
    Doing random r/w test
    Initializing worker threads...

    Threads started!

    File operations:
        reads/s:                      46421.95
        writes/s:                     30947.96
        fsyncs/s:                     99034.84

    Throughput:
        read, MiB/s:                  725.34
        written, MiB/s:               483.56

    General statistics:
        total time:                          60.0005s
        total number of events:              10584397

    Latency (ms):
             min:                                    0.00
             avg:                                    0.01
             max:                                    1.32
             95th percentile:                        0.03
             sum:                                58687.09

    Threads fairness:
        events (avg/stddev):           10584397.0000/0.00
        execution time (avg/stddev):   58.6871/0.00

2TB Nand Samsung 980 Pro: Prep time: `sysbench fileio --file-total-size=150G prepare` 161061273600 bytes written in 87.15 seconds (1762.53 MiB/sec).

    Benchmark:
    `sysbench fileio --file-total-size=150G --file-test-mode=rndrw --max-time=60 --max-requests=0 run`
    WARNING: --max-time is deprecated, use --time instead
    sysbench 1.0.20 (using system LuaJIT 2.1.1741730670)

    Running the test with following options:
    Number of threads: 1
    Initializing random number generator from current time

    Extra file open flags: (none)
    128 files, 1.1719GiB each
    150GiB total file size
    Block size 16KiB
    Number of IO requests: 0
    Read/Write ratio for combined random IO test: 1.50
    Periodic FSYNC enabled, calling fsync() each 100 requests.
    Calling fsync() at the end of test, Enabled.
    Using synchronous I/O mode
    Doing random r/w test
    Initializing worker threads...

    Threads started!

    File operations:
        reads/s:                      594.34
        writes/s:                     396.23
        fsyncs/s:                     1268.87

    Throughput:
        read, MiB/s:                  9.29
        written, MiB/s:               6.19

    General statistics:
        total time:                          60.0662s
        total number of events:              135589

    Latency (ms):
             min:                                    0.00
             avg:                                    0.44
             max:                                   15.35
             95th percentile:                        1.73
             sum:                                59972.76

    Threads fairness:
        events (avg/stddev):           135589.0000/0.00
        execution time (avg/stddev):   59.9728/0.00

4TB Sabrent Rocket 4 Plus: Prep time: `sysbench fileio --file-total-size=300G prepare` 322122547200 bytes written in 152.39 seconds (2015.92 MiB/sec).

    Benchmark:
    `sysbench fileio --file-total-size=300G --file-test-mode=rndrw --max-time=60 --max-requests=0 run`
    WARNING: --max-time is deprecated, use --time instead
    sysbench 1.0.20 (using system LuaJIT 2.1.1741730670)

    Running the test with following options:
    Number of threads: 1
    Initializing random number generator from current time

    Extra file open flags: (none)
    128 files, 2.3438GiB each
    300GiB total file size
    Block size 16KiB
    Number of IO requests: 0
    Read/Write ratio for combined random IO test: 1.50
    Periodic FSYNC enabled, calling fsync() each 100 requests.
    Calling fsync() at the end of test, Enabled.
    Using synchronous I/O mode
    Doing random r/w test
    Initializing worker threads...

    Threads started!

    File operations:
        reads/s:                      2690.28
        writes/s:                     1793.52
        fsyncs/s:                     5740.92

    Throughput:
        read, MiB/s:                  42.04
        written, MiB/s:               28.02

    General statistics:
        total time:                          60.0155s
        total number of events:              613520

    Latency (ms):
             min:                                    0.00
             avg:                                    0.10
             max:                                    8.22
             95th percentile:                        0.32
             sum:                                59887.69

    Threads fairness:
        events (avg/stddev):           613520.0000/0.00
        execution time (avg/stddev):   59.8877/0.00

dmayle

6 месяцев назадdiscuss

Fun...

This is something I have been thinking about and researching for awhile, because there is so very much confusing language out there.

Your quote says over the last century, so I'm going to use roughly 1920 as the baseline. It also refers to a per capita increase of meat consumption by 100 pounds, or about 45.4 kilograms (to make the math easier). This is roughly an increase of 124g of meat per person per day (or about 3oz if that makes more sense to you).

This equates to a daily increase in per-capita protein intake by 25-30g (depending on which meat and how lean it is).

In 1920, the average American adult male was about 140 pounds, and ate about 100g of protein per day, which works out to roughly 0.71 grams per pound of body weight (or about 1.6 grams per kilogram).

In 2025, one century later, the average American adult male is 200 pounds, and if he eats the same ratio of weight to protein, you would expect that he would eat around 140g of protein per day, which is slightly higher than the increase in per-capita meat consumption over the same time.

However, if you look at actual statistics of what people are eating in protein, you'll see that the average American adult male is actually eating about 97g of protein per day, or about 0.49 grams per pound (1.1 grams per kg), which is much less than we ate a century ago, which means that that the increase in meat consumption doesn't match change in protein, so is offset by either less non-meat protein, meat with lower protein content (e.g. more fat), or both.

There was some discussion lower in the thread about bodybuilders vs normal people, and about basing your calculations on lean body weight vs full bodyweight. Lean body weight calculations are often used for bodybuilders, but those numbers are elevated (typically 1 gram of protein per pound of lean body weight). For someone who is sedentary to lightly active (e.g. daily walks), the calculation is based on full body weight, not lean body weight, and is about 0.7 gram per pound (or 1.5 grams per kilogram), which matches this recommendation exactly.

Hitting these targets has been shown to greatly increase satiation, reduce appetite, but it does not make you lose weight, and it is not permanent (reducing your protein intake removes the effect, which makes sense). However, long term studies show that people who increase their protein intake to these levels and lose weight (through calorie reduction or fasting) keep that weight off.

Finally, from what I've been able to cobble together, high protein intakes combined with high fat and high sugar intakes does not have the same effect as a diet that matches the recommendations here (ie. it's not just about higher protein intake, it's about percentage of calories from protein, which should be around 20-25%... 200 pound sedentary to lightly active adult male, 140g of protein, or 560 calories, in a total diet of 2250-2800 calories, depending on activity level)

dmayle

7 месяцев назадdiscuss

The sad part of all of this was that the company that does this tried to poach me back in 2013 or 2014, but I was disgusted by the practice, so I refused to even interview.

Since then, I've made sure every single TV I own has this turned off (I go through the menu extensively to disable, and search on Google and reddit if it's not obvious how to disable like the case with Samsung).

I have an LG Smart TV, and just a week or two ago I was going through the settings and found Live Plus enabled, which means either they renamed the setting (and defaulted this to on), or the overrode my original setting.

Either way, I'm super annoyed. I want to switch to firewalling the TV and preventing any updates, but I need a replacement streaming device to connect to it.

Does anyone have recommendations for a streaming device to use (presumably one with HDMI CEC, that supports 4k and HDR)? I use the major streaming services (Netflix, Prime, Hulu, Apple TV) and Jellyfin.

dmayle

7 месяцев назадdiscuss

I used to manage a team working on the news feed at Facebook (main page).

We did extensive experimentation, and later user studies to find out that there are roughly three classes of people:

1) Those that use interface items with text 2) Those that use interface items with icons 3) Those that use interface items with both text and icons.

I forget details on the user research, but the mental model I walked away with this that these items increase "legibility" for people, and by leaving either off, you make that element harder to use.

If you want an interface that is truly usable, you should strive to use both wherever possible, and ideally when not, try to save in ways that reduce the mental load less (e.g. grouping interface by theme, and cutting elements from only some of the elements in that theme, to so that some of the extra "legibility" carries over from other elements in the group)

dmayle

7 месяцев назадdiscuss

I recently learned about the (ancient?) greek concept of amathia. It's a willful ignorance, often cultivated as a preference for identity and ego over learning. It's not about a lack of intelligence, but rather a willful pattern of subverting learning in favor of cult and ideology.

dmayle

7 месяцев назадdiscuss

The only actual problem with cheating is leaderboards.

When you have accurate matchmaking, you will be playing against other players of a similar skill level. If you we're playing in single-player mode, it wouldn't bother you that some of the players were better than others.

Whether the person you're playing against is as good as you because they have aim assist, while you have a 17g mouse and twitch reflexes shouldn't matter. You're both playing at equivalent skill levels.

The only reason it matters to anyone is that they want their skills to be recognized as better than someone else's. Take down the leaderboards, and bring back the fun.

I say, let the people cheat.

dmayle

8 месяцев назадdiscuss

I've been using an 8k 65" TV as a monitor for four years now. When I bought it, you could buy the Samsung QN700B 55" 8k, but at the time it was 50% more than the 65" I bought (TCL).

I wish the 55" 8k TVs still existed (or that the announced 55" 8k monitors were ever shipped). I make do with 65", but it's just a tad too large. I would never switch back to 4k, however.

dmayle

10 месяцев назадdiscuss

Definitely not useless!

I run a ttyd server to get terminal over https, and I have used carbonyl over that to get work done. That's limited to a web browser (to get access to resources not exposed via the public internet), so having full GUI support is very useful

dmayle

10 месяцев назадdiscuss

I am truly shocked by this.

I looked it up, and it turns out you're right. Both the iPhone 17 and the iPhone Air use USB2.

USB3 was introduced in 2008 (!!!). That is 17 years ago.

I already wasn't interested in this tech, to be fair, but I've had to support family phones synchronizing/backing up over the cable, and even at full theoretical speed for the transfer, we're talking over an hour vs just under 7 minutes. Which, considering the flash most likely suppports the read in under a minute, is crazy.

dmayle

10 месяцев назадdiscuss

Which is literally what Apple announced in this video:

"and the 2x telephoto has an updated photonic engine, which now uses machine learning to capture the lifelike details of her hair and the vibrant color of her jacket"

"like the 2x telephoto, the 8x also utilizes the updated photonic engine, which integrates machine learning into even more parts of the image pipeline. we apply deep learning models for demosaicing"

dmayle

11 месяцев назадdiscuss

I used to work at Meta (back when it was just Facebook), and I pioneered a similar effort back in 2016-2017-ish. Now, I don't know anything about the current version (which seems to offer cloud processing as well), but when I was there, the effort was entirely local to the phone.

We had caffe2 running a small model on the phone to try and select and propose photos for the user to share.

We were trying to offer an alternative sharing model that both made sharing easier, while offering the user the controls that made them feel comfortable with photo suggestions. (for those who never noticed, we launched Moments, which was an app that allowed automatic private sharing of your camera roll with a close selection of friends and family, but the experience wasn't great because it was centered around group events and sharing photos with the people who were there, not connecting with the ones who weren't)

Ultimately, it was scrapped, because we were paranoid that we hadn't come up with a user experience that made it clear that this was happening only on the phone (I think we even tried a notification model), or that we'd accidentally surface someone's boudoir photos, and we were too worried about the kind of knee-jerk reactions that you're seeing in this thread.

I'm guessing that someone at Meta either had a more successful go at the UX, or they feel that the opinions about AI have shifted enough that there will be less fear.

Upon reading the article, it looks like there are two options, one which is local-only, and similar to what we built, and a second one which tries to make better suggestions using online, and that is only enabled after asking the user.

I would suspect that the cloud processing version also runs a local model to attempt to filter out racy photos before sending them to the cloud, but I don't know for sure.

I think the article is a bit disingenuous in it's presentation, but it's possible that I'm biased because I know how a similar thing was built, but it definitely sounds like fear-mongering.

Submissions

[untitled]

comments