Hacker Newsnew | past | comments | ask | show | jobs | submit | more l5870uoo9y's commentslogin

Long-term memory storage capacity[1] scores 0 for both GPT-4 and GPT-5. Are there any workable ideas or concepts for solving this?

[1]: The capability to continually learn new information (associative, meaningful, and verbatim). (from the publication)


Static content resolves correctly but data fetching is still not functional.


It is appreciated that you can change the temperature unit by clicking on it, and how surprisingly cold and changeable the temperature is as you travel up through the atomic sphere (down to -84C, -119F).


Inflation rose by 6.5 cents/year in the period 2020-2022, and if that were the case for the entire period, the 50 cents would have grown to 201.5 cents.


I am in the process of porting my web app[1] from NextJS hosted at Vercel to Astro/React hosted at Cloudflare, and something that particularly surprises me is that I can render a web app on every request at “the edge” and have response times of 100-200ms. That is almost as fast as fully static pages.

I have also definitely noticed an improvement in Cloudflare Worker over the last few weeks; cold starts have practically disappeared, and they are significantly more stable in terms of response times.

[1]: https://app.sqlai.ai/


Hello! How are you collecting your edge workload to the database? Are you using cloudflare’s database?

I’ve wanted to try out this edge hosting thing but because of the amount of round trips involved between the application and the database, the application performs worse on the edge.

Thanks!


The database is hosted on Railway. I have enabled Smart Placement that should, depending on request time, automatically use an endpoint closer to the database to speed it up. When requesting a route that includes a database request, the response time on the US east coast is around 200ms and closer to 1000ms in Denmark. I am hoping that the Smart Placement will work better when I go live with the app (still in beta) and that it mainly needs more traffic to calculate optimal endpoint placement.


Have you given Hyperdrive[0] a try? In theory, it should improve performance in your use case where you have a central database (Railway) being connected to from the Edge (your Workers).

It moves the DB connection logic closer to your Workers, pools connections, and can also cache queries.

(Disclaimer: I work for Cloudflare, but on an unrelated team. Not personally used Hyperdrive, but heard good things!)

[0]: https://developers.cloudflare.com/hyperdrive/


It actually looks very interesting, but my app is not particularly affected by database query time, but rather by the time spent on the various AI generations.


So a remote, authenticated DB connection over the internet? How many db queries per page generation does that generally work out as?


My app isn't database heavy; most pages renders without a single database call (user information stored in cookie) and the pages that do call the database contain maximum one call. Other data, like database schemas and subscription status, are requested in the background directly from the browser so it doesn't stall page rendering.


C'mon - static pages are like 10ms or less. 200ms is already noticeable, not-instant for humans. We have 5-10x faster hardware and 10x slower websites ;)


What is the smallest fraction of a Nobel Prize that a winner can be awarded?


1/4. The max number of recipients of any prize is 3, but it can be split into a half, and then one of those halves into two further halves, for a 1/2 + 1/4 + 1/4 split. (It can also be split equally into 1/3 + 1/3 + 1/3).



Everyone talks about Alzheimer's and dementia, but Daniel Kahneman has neither. He chose to commit suicide because he wanted to avoid “natural decline.” That's an unexpected statement from a 90-year-old. I'm more surprised by his lack of will to live and that he just “gives up” and throws away the most valuable thing he has.


I think it's more the Nobel Peace Prize that stands out.


The peace prize stands out for being different.

The economics one stands out for not being endowed by Nobel but instead Sveriges Riksbank well after his death (thus it's the Nobel Memorial Prize in Economic Sciences).

But it's administered by the Nobel Foundation, etc.


The Peace Prize stands out because sometimes it is given to someone who wages war such as Henry Kissinger or Barack Obama.


Giving the prize to Obama, whose main achievement at the time was not to be Bush, was a disgrace and did huge damage to the image of the committee. And he did nothing after that that would have deserved a prize.


Bush's PEPFAR was so good that if they gave him the prize it'd be more deserved than many other people who have it.


I must admit that I disagree and believe that AI will only be used more as it improves. There is already a big difference between GPT-4 and GPT-5 in terms of hallucinations. If AI can do your tax accounting in 20 seconds with a 95% probability of accuracy (perhaps better than most accountants—after all, who can really understand tax legislation?) and with some clear checksums, who wouldn't use it instead of doing it themselves? In addition, AI is really excellent at inferring meaning from data (and see relations).


For most folks, filing your taxes is already mostly automated and requires a time investment of maybe 45 to 60 minutes per year. Would it be nice to reduce that down to 20 seconds? Sure, but it's not going to materially change anything in the financial life of most people. I'm all for it if it could be done reliably at scale. Outlier cases where taxation is complicated enough that it's possible to engineer substantial and meaningful financial outcomes often count more as business finance than personal finance. I'm sure artificial intelligence will indeed be more disruptive to business and corporate finance. For true personal finance, though, I suspect it may be more predatory than helpful by selling people into bad decisions that sound smart (which banks, etc., already try to do with so-called financial advisors).

Keep in mind a 5% failure rate would likely be >10 million incorrect tax filings per year solely due to AI errors, not inclusive of additional incorrect filings due to human error as well.


Better idea: the U.S. could adopt the income tax practices of most other countries, where the revenue service tells you what you owe, instead of making you guess. No need for A.I. or Intuit.


Fwiw iirc Intuit has to offer a free version of their tool with the same capabilties but it is so hidden from everybody behind so much stuff that its hard for people to reach there.

Honestly, yes there shouldn't really even be a discussion about this. US should definitely do it. I think US and India are the only two major countries still stuck on something like this and I am not sure if there are really any advantages of it.


They will tell you how much you owe, but only if you file incorrectly :P

For Americans, if you didn't know, you can check your IRS transcripts at https://www.irs.gov/ to see exactly how much the IRS thinks you've earned.


Thank you, I didn't know about this and it's great information to have for reference!


That doesn't arbitrarily increase GDP through revenue-through-friction, though.


> perhaps

That perhaps is doing a lot of lifting.


Not really. After using GPT-5 extensively I'd easily trust it to do my taxes. I find the output to be very reliable, also my taxes are trivial.


I'm currently doing that. One problem is that it's interpretation of the tax law and instructions are not consistent. My taxes are not trivial.


> I'd easily trust it to do my taxes. I find the output to be very reliable, also my taxes are trivial.

Tax season in the US is in about six months - I look forward to hearing how your experiment goes!


For the 95% of people who get all their income from W2/1099 compensation, bank accounts, and brokerage accounts and who either take the standard deduction or whose only itemised deductions are dependents, mortgage interest, and SALT, there's really no need for filing tax returns at all. But the tax accounting and tax software industries lobby to prevent it.


In the current AI = LLM world, why have a language model do taxes? Why not just have AI help you adapt your non AI tax planning platform to local tax laws by reading and comprehending them at scale, a language task, instead?


5% odds of screwing up my taxes and going to jail through no fault of my own is a horrible proposition.


Unrelated question, where are these cores produced?


TSMC.


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

Search: