Hacker News new | past | comments | ask | show | jobs | submit | ashirviskas's comments login

Why not? There's still 7 months left for breakthroughs.


Small leaves wiggle room, but it's extremely unlikely trad small, <= 7B, will get there this year even on these evals.

UX matching is a whole different matter and needs a lot of work: Worked heavily with Llama 8B over last days, and Phi 3 today, and the Q+A benchmarks don't tell the full story. Ex. It's nigh impossible to get Llama _70_B to answer in JSON; when Phi sees RAG from search it goes off inventing new RAG material and a new question.


Both??


I've read that Matthew Walker isn't a good source on sleep, as he tends to just make up some claims or bend the data to support whatever point he wants to make [1].

[1] - https://guzey.com/books/why-we-sleep/

EDIT:

>Melatonin levels peek in the middle of the day and are lowest around midnight.

Isn't that the opposite of how melatonin works?


Based on https://slatestarcodex.com/2018/07/10/melatonin-much-more-th..., as well as every chart of melatonin I can find, it doesn't seem to peak during the day. It seems to rise before bedtime and peak in the early hours of sleep.

According to the blog post, the treatment for some sleep disorders is to take a dose of melatonin during the day, which moves your natural melatonin production earlier for unclear reasons. Maybe there's confusion between when you take the melatonin and when melatonin is naturally highest?


Yeah, it was my understanding that melatonin is secreted under low-light conditions according to one's circadian rhythm.

> A substantial number of studies have shown that, within this rhythmic profile, the onset of melatonin secretion under dim light conditions (the dim light melatonin onset or DLMO) is the single most accurate marker for assessing the circadian pacemaker. Additionally, melatonin onset has been used clinically to evaluate problems related to the onset or offset of sleep. DLMO is useful for determining whether an individual is entrained (synchronized) to a 24-h light/dark (LD) cycle or is in a free-running state. DLMO is also useful for assessing phase delays or advances of rhythms in entrained individuals. Additionally, it has become an important tool for psychiatric diagnosis, its use being recommended for phase typing in patients suffering from sleep and mood disorders. More recently, DLMO has also been used to assess the chronobiological features of seasonal affective disorder (SAD). DLMO marker is also useful for identifying optimal application times for therapies such as bright light or exogenous melatonin treatment. [1]

[1] https://pubmed.ncbi.nlm.nih.gov/16884842/


>Flops really are quite cheap by now, e.g. vision inference chip ~$2/teraflop/s !!

I'm really interested, can you share where you got these numbers?


Axelera [1] or Halio [2] give you 100-200tflop for ~$200.

8-bit ops, inference only, low memory embedded, excluding the host, implied utilization from FPS specs is ~20%

But the trend is there.

There are also newer ADAS/AV units from China which claim 1000tflops and cant really cost more than $1000/$2000 per car.

These are all tiled designed (see also dojo/tesla) heavily over-weighed on flops vs memory

[1] https://www.axelera.ai/

[2] https://hailo.ai/


You can't get flops on a Hailo-8, they're fixed-point only. As much as these specialised inference chips are cool, we're a long way from just being able to drop them in where a GPU was. Not to mention the memory is hugely constrained. The Hailo chips I've worked with were all limited to 20MiB for the weights which is a squeeze even at 4-bit.


Yeah, and Belarus is doing so much better than those "degrading" baltics.


I currently have OpenAI subscription and I am really interested in moving to Kagi, but it is hard to find information on their ultimate plan. What AI features does it offer?


You can choose between gpt-3.5, gpt-4, claude-1, claude-2, and google 'bison' (I didn't try that one though).

Mysteriously, it seems like the chat box on openai.com gave me slightly better answers than via their api; and sometimes Kagi.com renders a bit poorly while streaming responses, but meh. I can throw a random istio error into it or whatever and it gets me 75% of the way to a solution, so it works well enough. YMMV


Thanks! Can you do image generation, or upload your own images for it to use?


> I have ptsd just thinking about [...] what it means to have hardware devices with drivers that barely work most of the time.

Same, but for my macbook pro. Bluetooth implementation is borked and I can't even use a single device properly - I have to constantly unpair/pair my headphones. If I use 2 devices there is some stuttering in sound. But if I use 3 - all devices stutter, unless I kill some bluetooth daemon, then it works normaly for 2-5 seconds and goes into the bork mode again.

But then the software is awful too:

1. Switching between workspaces takes up to a full second before the new workspace becomes active even after disabling all animations

2. Some windows just randomly decide to not show the three control buttons and it becomes impossible to close them without messing with the process.

3. For simple screen recording I have to open Quicktime player. Then the screenshot tool becomes screen recording tool with no apparent way to return it back to the screenshot tool.

And these are just the ones I experienced this/last week. Don't get me started on mouse getting stuck in a secondary monitor or disappearing completely and other shitty UX experiences I've had being forced to work with macos for the past 2 years. Can't wait to move away and not look back.


Can't comment on the Bluetooth issue, since Apple devices have the least BT issues for me by far. But for screen recording there's the cmd+shift+5 shortcut (or the Screenshot App) to do the same thing without opening Quicktime player.


It is the only platform giving these bluetooth headaches and the reason my nice BT headset sat in the closet for almost a year.

Regarding the screenshots, I've just learned these shortcuts, but if I open the screenshot tool using CMD + Spacebar + "screenshot", it often just defaults to screen recording and I see no option to switch that.


>what? people are following direction, and the way the plane boards is actually important to how fast it loads, so youre messing with that because its a mild convenience to you and you consider that intelligent behaviour?

What direction? The only time there were more directions than just "priority" vs "non priority" was on my flight to Korea a few years ago. Most smaller flights in Europe, sadly, do not do that.


Full service airlines board by section by direction of the gate agents pre-flight. The video explains why its important for those airlines


I've watched that video when it came out, but my point is that you can fly monthly between european countries without ever coming across proper boarding procedures. Don't get me wrong, it infuriates me, but this is the sad reality we live in.

Even on the flight from Helsinki to Bangkok there was only priority vs non-priority distinction, even though the tickets had different seating groups. Could have saved us all some convenience, but no, we all had to board the plane very chaotically.

Which is why I am usually the last one to go to the boarding, I can chill on comfy seat for 20-30 minutes instead of standing in chilly or boiling hot weather/boarding tunnels.


I just found out my country also started deprecating 3G last year and will deprecate 2G in 2025. That is a bit sad, as I've just powered on some old 2G phones and they seemed to work just fine, but I guess it allows us to move forward without the tech debt.


Tech debt in this case seems like artificial scarcity


Lack of spectrum and/or tower sites is real scarcity.

LTE and 5G are supposed to be better able to share spectrum, to the point where they can share the control channel, and multiplex the data channels with either encoding, but 2g and 3g can't do that, so you have to dedicate at least minimum sized blocks.

Running one min block on 2G does wonderful things for ensuring access, but it's expensive in spectrum, especially if each network needs to do it. (Because cooperating between networks is really only a thing if mandated, or out in rural areas where there's little demand)


Do you still use dial up?


I use both Slack and Discord webapps in the according profile, why not use that? They already run in sandboxed browser instances (I guess), so there's even less overhead running them in your browser. (I did not do extensive testing, so I might be wrong here)


You are basically saying "Don't use any native apps, ever". With chrome profiles, links are just opened in whatever profile window was last active (this is on Mac OS, but I think others are the same). It's the one thing keeping me with chrome (brave).


Linux is definitely not the same. The browser instance that was opened first wins, unless you start it with `--no-remote` flag.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: