It says that the company is co-founded by Postgres creator. I find that bit hard to believe given that there is nothing novel in the article, probably discovery for them. I do understand that everyone has to go through their own journey to learn these things but at the same time when you are running business then seeking professional help isnt a bad idea.
Based on my experience queries like these cannot scale, whatever you do. However if you are already on a path where you had invested a lot in such queries then hire a DBA, if you are not far off then hire an architect to model the data for better performance.
My understanding (possibly flawed) is that Stonebraker isn't directly involved anymore.
The original idea that he worked on with the DBOS people at MIT and Stanford was very different and much, much more ambitious, which is why the name DBOS seems a little out of place now. The original idea was much closer to a "database OS".
Here [1] is the paper, which proposes that "To improve the scalability, security and operability of OSes, we propose a data-centric architecture: designing the OS to explicitly separate data from computation, and centralize all state in the OS into a uniform data model. In particular, we propose using database tables, a simple data model that has been used and optimized for decades, to represent OS state. With the data-centric approach, the process table, scheduler state, flow tables, permissions tables, etc all become database tables in the OS kernel, allowing the system to offer a uniform interface for querying this state."
The team later published another paper based on their prototype work [2].
Instead, they basically implemented Temporal as a client library with Postgres as the state layer. It's good, but only tangentially related to the original vision.
Maybe the long-term plan is an actual database OS, but it kind of looks like they decided they had to pivot to something much simpler, and slapped on an "for AI" like everyone is doing these days.
Scale with what? If you have m distinct values in an index, then listing them this way takes m log(n) time, which is fine for many use cases no matter how much data you have.
The way the OP is trying to achieve all the goals by pushing the complexity on the queries/database is what I am referring to as non-scalable as data grows on SQL DB.
> If you have m distinct values in an index, then listing them this way takes m log(n) time, which is fine for many use cases no matter how much data you have.
And NO the runtimes are not right away applicable on machines at scale. You are dealing with DB locks, page sizes, available memory, existing data in memory, queue depth. Experienced folks get paid to short circuit such learnings
Runtimes are usually pretty well applicable at scale, it’s just that most people don’t have a good intuition about asymptomatic notation. Constants and lower order terms matter a lot in practice but are hidden in asymptotic notation.
First of all, there is definitely value addition with the LLMs in almost every field in some ways.
What bothers me the the marketing angle which invites skepticism and criticism
> Anthropic invited Matt von Hippel to write this post and compensated him for his time.
If you are paying some one, tailoring the discussions then the end result is always going to be biased one showing yourself as the winner. I understand its somewhat organic, still the ratio of marketing and science needs to be balanced. Marketing has to be correct and the results/outcomes should be reproducible
AI has its own benefits but its really a bad idea to use it in education as a primary source. LLMs have lots of information, reasoning ability, however they have their own limitations and by the nature these limitations just amplify.
Few days ago my son was working on a problem which asked if square is a rectangle. Turns out his class teacher and external tutor (who was using AI) had different opinions on it. My wife consulted AI as well and she also came to the conclusion that square is not a rectangle cause AI told her so, AI told the same answer to external tutor. I understand that the context matters and if one probes AI more with additional context it may come up with correct answer, but this depends on the assumption that the person asking AI for something already knows the answer.
This is not a problem with AI, this is a problem with a human accepting an answer without understanding why it is correct. The tutor could also read a book that states "A square is not a rectangle" (or vice versa) and repeating it back to the student as established fact.
The whole point of having teachers is to give intellectually developing and still vulnerable children the best possible support they could have from a human which is ideally interested in their well-being, being responsible and in doing a good job.
Replacing that with AI is the kind of SV-stupid which starts making sense after the educational system was hollowed out for profit reasons and is failing. The human in that system is just checking in and DGAF.
I am very curious what either person asked that led them to say that AI had told them a square is not a rectangle.
The obvious question of “is a square a rectangle?” is definitely NOT causing any LLM to say “no”.
Hallucinations are an issue but they genuinely are no longer a concern for any modern non-tiny LLM for basic facts that you would expect a K-6 course to cover. Maybe you might be able to find a few random hallucinations in knowledge covered at the high school level but I’m quite skeptical even that still happens. Yes, they can still, at times, hallucinate sources but I cannot recall the last time I actually saw this happen. I think it was at least a year ago
I read their comment as the teacher (who was correct) and a tutor (who used AI) disagreed, not two tutors.
With the amount of blatantly false information I've seen LLMs spew out in response to my very simple questions, I can very easily imagine one would get this basic information wrong somehow, although I did just Google it (as Gemini, and particularly the on-search-result-page version of it, seem to be the absolute worst of the common models in my experience), and it did get the answer right.
When shape names are first introduced, many kids will flip out if you suggest a given polygon can be multiple things at once. So they define the shapes (square, rectangle, rhombus, parallelogram, trapezoid, etc.) so they are mutually exclusive, with categories like quadrilateral as residual categories. You can tell who is doing AoPS/Beast Academy because they argue based on the standard definitions, whereas the teacher will insist on the curricular definition or on the “most specific answer.”
Another classic is “how many roots does a parabola with real coefficients have?” The answer is always two by the FTA, but the textbook Algebra 1 answer is zero, one, or two because they haven’t covered complex numbers yet.
Let me say this in a different way: If some rectangles can also be squares, then this shows that a square is a rectangle. This is not a logical deduction; it is an intuitive one.
LLMs have lots of information, but they have zero reasoning ability. That is why they are so bad at giving answers: they can't reason from one thing to another, but instead everything is a statistically likely result. If using an LLM for learning, you have to make sure to check everything it says or you will take in false info, guaranteed.
LLMs have had a reasoning setting for a while now. The reasoning is what allows them to go through and test each output to see which output/answer ends up aligning with known context the best (whether that be constraints that steer it away from some outputs, or information/hints that steer it towards the desired output).
The problem with the tutor's question is that they're going to AI to answer this question and not asking underlying questions like "what makes a square a square" and "what makes a rectangle a rectangle", then doing their own thinking to determine if a square's qualifications also fit inside the qualifications to be a rectangle.
Heck, even if you just feed the underlying into AI it'll reason the answer in a logical manner:
> In standard geometry, A square is a quadrilateral with 4 equal sides and 4 right angles (90° each).
A rectangle is a quadrilateral with 4 right angles. Its opposite sides are equal and parallel, but all four sides do not have to be equal. So the key distinction is: Rectangle: four right angles. Square: four right angles plus four equal sides.
>
> That means every square is a rectangle, but not every rectangle is a square.
Asking LLMs to just give you answers is the real antipattern. But if you remember that the only thing it's good at is finding patterns given context/constraints, it can be an extremely good tool.
Also, I have a feeling the vast majority of the populus is using free versions of Claude or ChatGPT. Thankfully recently ChatGPT Free is GPT-5.6 Luna, but before the response qualities and capabilities of the free chat output were less than stellar.
Wait a second, if you decided to become intermediary in the protocol then you are supposed to add value, not take away the features that already exists.
Although its not clear from the article itself, my gut feeling is that some big enough client arm twisted them to support it before they sign the contract again
Implementing the HTTP/1.1 caching mechanism at the proxy server is delightfully simple if you just ignore it entirely. But if you decide to do some caching on your own, you better implement the semantics in the way the clients and the servers expect it to be. Which is not simple at all if you're doing the caching to eke some performance improvements.
I have a similar to pagerank method I use to evaluate such papers. I look for the references to see how many authors are using their own references (past work), the idea being that people do not jump too far, they make incremental progress.
For the pre-print I could only find only one author who has a single referenced article.
> references to see how many authors are using their own references (past work), the idea being that people do not jump too far, they make incremental progress
The authors are not using their own prior work in the paper, thats the point I was trying to make. I have worked in biotech lab for couple years and its one of the criteria's people use to consider some ones work useful and worth the time.
I think you fundamentally misunderstand both the results published here, and how scientists operating at the highest level of academic research operate. None of these authors has to worry about citing previous work to get the attention of biotech labs.
Its beneficial for the company for sure at this stage cause most of the people using these LLM tools and agents have built the muscle to be able to review the code so the work gets done faster. These tools are definitely helpful in lot of ways to move faster.
Where it does not helps, software engineering is a field where you have to actively practice it to make sure your mind remembers the trade offs, different paths you have evaluated before, sort of building muscle memory. Using LLMs mean that eventually your debugging skills, ability to understand some one else code would degrade and moving to the next level in terms of software competency may never happen for you.
A concrete example, the ability to remember 10 digit phone numbers was really normal before smart phone, try to push your mind now to see if you are still able to do that?
Lastly, at any given point of time 90% of the engineers were working on managing the existing code base, so career wise with so much code being written the job safety should not be an issue if you have skills to understand and debug some one else code
> A concrete example, the ability to remember 10 digit phone numbers was really normal before smart phone, try to push your mind now to see if you are still able to do that?
But would you tell anyone in the early 2000s to keep practising 10 digit phone number memorisation knowing that it will be completely useless skill in the coming future? How do you know what telling OP to be more hands on and spend time on active practice isn't the same thing?
Its a good question for thought, being able to remember a number is a mind capability, same applies for maps too as another concrete example. I know so many people around me who use maps even for the same route that they have been traveling for months.
Doesn't happens that often, though I have seen them struggling when maps are unavailable, same for people struggling to call some one in need if their phone battery is out or if they loose their phone.
May be I am in a minority here, but there are areas where SAML sort of shines
1. For OIDC/OAuth2 the request has to originate from Service provider, most Enterprise IDP's rely on SAML for Single Sign on cause of its ability to do IDP initiated flows
2. The security for SAML is baked into the payload itself, provides safety against MITM attacks, even though the HTTPS provides similar guarantees in theory, the reality is that your SSL offloading happens elsewhere, not on your application server
The OP provided a list of vulnerabilities discovered in SAML, IMO this kind of comparison if flawed if you do not present the same for OIDC/OAuth2.
At the end of the day, these are tools and effectiveness of a tool is a lot dependent on ones skillset to understand and use the tool.
> The security for SAML is baked into the payload itself, provides safety against MITM attacks, even though the HTTPS provides similar guarantees in theory, the reality is that your SSL offloading happens elsewhere, not on your application server
This is just as true for JWT. Actually its a lot more true for JWT as people usually implement this wrong in SAML.
Additionally, https doesn't really protect you here, as typically the user sees the token/saml document, and they are the main party you have to worry about being in the middle.
> this kind of comparison if flawed if you do not present the same for OIDC/OAuth2.
i'm doubtful there are any vulns in oidc/oauth2 that aren't present in saml. SAML vulns are typically a strict superset of oidc vulns.
No, most workers don't really say this in surveys and polling. They tend to hate LLM tools because it makes their jobs worse, nothing about being left behind.
The only people pushing the "left behind" narrative is SV + SF + VC since their previous narratives have failed to persuade the public (thank fuck).
well people are spending tokens like crazy for sure, Anthropic and all other labs and everything related to AI seem to be making tons of money.
On a serious note, it might take a while to realize the actual benefits or losses. Its clearly not a good signal when people whose job is to manage other people start writing their own pet AI projects. This just indicates that AI has created this big job insecurity among everyone. At the end of the day being an engineer, eventually the job is relatively safer when there is so much code being written out there.
Yes the driver made a poor decision in a situation that he was not trained for. However enrolling and onboarding the driver is a part of Uber's process. The situation may have been avoided if Uber had this strict clause in its onboarding manual to never stop the car or drop the passenger at unsafe spot. The driver and rider are bound by the Uber negotiated contract atleast until the end of ride.
On the driverless cars, the chances are that the companies would come up with some innovative legal clause to add to their terms and services to avoid taking any liability from the driverless car's action.
> The situation may have been avoided if Uber had this strict clause in its onboarding manual to never stop the car or drop the passenger at unsafe spot.
Its interesting, saying sorry for your mistakes using plain words does not have the same impact as saying it in a convincing manner as if you mean it. Giving them a 10 page long documentation and asking them to just acknowledge it as a part of onboarding process wont do the job. We are talking about human life here and an emphasis on security should be baked into mind by regular training, and it is usually required for such jobs in general.
Being responsible for some one else's safety should have a higher bar than being responsible for your own safety.
I cannot claim to have done the onboarding process myself, tried to look up online and the most I could find is that a good amount of time is spend on doing background checks. I was able to find Uber's employee onboarding information for their corporate employees and it turns out that there is indeed some training required for people who are managing your corporate property. Ironically I do not see any mandatory (ideally paid for) training provided by Uber to drivers when they are handling other human beings (passengers). And to be fair its hard to predict everything, mistakes happen unfortunately, the goal should always be to get to a better state than where we are as of today.
Based on my experience queries like these cannot scale, whatever you do. However if you are already on a path where you had invested a lot in such queries then hire a DBA, if you are not far off then hire an architect to model the data for better performance.
reply