Actually, this has already happened in a very literal way. Back in 2022, Google DeepMind used an AI called AlphaTensor to "play" a game where the goal was to find a faster way to multiply matrices, the fundamental math that powers all AI.
To understand how big this is, you have to look at the numbers:
The Naive Method: This is what most people learn in school. To multiply two 4x4 matrices, you need 64 multiplications.
The Human Record (1969): For over 50 years, the "gold standard" was Strassen’s algorithm, which used a clever trick to get it down to 49 multiplications.
The AI Discovery (2022): AlphaTensor beat the human record by finding a way to do it in just 47 steps.
The real "intelligence explosion" feedback loop happened even more recently with AlphaEvolve (2025). While the 2022 discovery only worked for specific "finite field" math (mostly used in cryptography), AlphaEvolve used Gemini to find a shortcut (48 steps) that works for the standard complex numbers AI actually uses for training.
Because matrix multiplication accounts for the vast majority of the work an AI does, Google used these AI-discovered shortcuts to optimize the kernels in Gemini itself.
It’s a literal cycle: the AI found a way to rewrite its own fundamental math to be more efficient, which then makes the next generation of AI faster and cheaper to build.
This is obviously cool, and I don't want to take away from that, but using a shortcut to make training a bit faster is qualitatively different from producing an AI which is actually more intelligent. The more intelligent AI can recursively produce a more intelligent one and so on, hence the explosion. If it's a bit faster to train but the same result then no explosion. It may be that finding efficiencies in our equations is low hanging fruit, but developing fundamentally better equations will prove impossible.
I was thinking more the numeric columns which have pre-built compression mechanisms to handle incrementing columns or long runs of identical values. For sure less total data than the text, but my prior is that the two should perform equivalently on the text, so the better compression on numbers should let duckdb pull ahead.
I had to run a test for myself, and using sqlite2duckdb (no research, first search hit), and using randomly picked shard 1636, the sqlite.gz was 4.9MB, but the duckdb.gz was 3.7MB.
The uncompressed sizes favor sqlite, which does not make sense to me, so not sure if duckdb keeps around more statistics information. Uncompressed sqlite 12.9MB, duckdb 15.5MB
I know I'm walking into a den of wolves here and will probably get buried in downvotes, but I have to disagree with the idea that using LLMs for writing breaks some social contract.
If you hand me a financial report, I expect you used Excel or a calculator. I don't feel cheated that you didn't do long division by hand to prove your understanding. Writing is no different. The value isn't in how much you sweated while producing it. The value is in how clear the final output is.
Human communication is lossy. I think X, I write X' (because I'm imperfect), you understand Y. This is where so many misunderstandings and workplace conflicts come from. People overestimate how clear they are. LLMs help reduce that gap. They remove ambiguity, clean up grammar, and strip away the accidental noise that gets in the way of the actual point.
Ultimately, outside of fiction and poetry, writing is data transmission. I don't need to know that the writer struggled with the text. I need to understand the point clearly, quickly, and without friction. Using a tool that delivers that is the highest form of respect for the reader.
I think the main problem is people using the tool badly and not producing concise material. If what they produced was really lean and correct it'd be great, but you grow a bit tired when you have to expend time reviewing and parsing long, winding and straight wrong PRs and messages from _people_ who have not put in the time.
While I understand the point you’re making, the idea that Excel is deterministic is not commonly shared among Excel experts. It’s all fun and games until it guesses that your 10th separator value, “SEP-10”, is a date.
The point made in the article was about social contract, not about efficacy. Basically if you use an llm in such a way that the reader detects the style, you lose the trust of the reader that you as the author rigorously understand what has been written, and the reader loses the incentive pay attention easily.
I would extend the argument further to say it applies to lots of human generated content as well. Especially sales and marketing information which similarly elicit very low trust.
I think often, though, people use LLMs as a substitute for thinking about what they want to express in a clear manner. The result is often a large document which locally looks reasonable and well written but overall doesn't communicate a coherant point because there wasn't one expressed to the LLM to begin with, and even a good human writer can only mind-read so much.
I’m with you, and further, I’d apply this (with some caveats) to images created by generative AI too.
I’ve come across a lot of people recently online expressing anger and revulsion at any images or artwork that have been created by genAI.
For relatively mundane purposes, like marketing materials, or diagrams, or the sort of images that would anyway be sourced from a low-cost image library, I don’t think there’s an inherent value to the “art”, and don’t see any problem with such things being created via genAI.
Possible consequences:
1) Yes, this will likely lead to loss/shifts in employment, but wasn’t progress ever like this? People have historically reacted strongly against many such shifts when advancing technology threatens some sector, but somehow we always figure it out and move on.
2) For genuine art, I suspect this will in time lead to a greater value being placed in demonstrably human-created originals. Related, there’s probably of money to be made by whoever can create a trusted system somehow capturing proof of human work, in a way that can’t be cheated or faked.
At this point, who really cares what the person who sees everything as "AI slop" thinks?
I would rather just interact with Gemini anyway. I don't need to read/listen to the "AI slop hunter" regurgitate their social media feed and NY Times headlines back to me like a bad language model.
In my master's thesis we used SUMO to model a small part of our town and hooked it up to the latest and greatest reinforcement learning algorithms to learn traffic light control. Eventually we beat all the other built in conventional algorithms in most parameters; Average speed. Emission. Etc.
Agreed! My team is constantly humbled by the mess of user data: names, birthdays, addresses, people dying or living abroad etc.
Honestly, sometimes I think about the linear algebra, AI, or robotics I learned in school and get this feeling of, "Is this what I'm doing? Stuff that feels like it should be simple?"
It's funny, even our product manager - who is a great guy - can fall into that "come on, this should be easy" mode, and I'll admit I sometimes get lulled into it too. But to his credit, every time I walk him through the actual edge cases, he totally gets it and admits it's easy to forget the on-the-ground complexity when you're in 'planning mode'.
So yeah, seeing your comment is incredibly validating.
Can someone explain to me how it makes sense that you want to define a locking mechanism using.. locking mechanism (the "atomic"). Does this mean that in an actual implementation you would have to drop down to some hardware-level-atomic operation, and is that even practical?
Also won't fencing token require some kind of token manager, that ensures you must present the highest token to do the action, and that you have to ask to get a new token, and that when you fail because your token is too old you must re-request one, is this modelled?
Well it's not a general purpose programming language defining the actual lock this way, it's checking a locking algorithm; so the atomic in the specification of the algorithm is just saying given that these steps do occur atomically, how will it behave.
The algorithm we're checking is using Redis, and the atomic read/write in the example is a behaviour Redis gives you.
I’m sure if I get hurt on my bike, my medical insurance coverage will not care at all whether the injury is from riding an e-bike.
If I hit somebody else on my bike… I don’t think I have insurance for that liability? I don’t think my auto or homeowner’s insurance policy mentions me riding bikes at all, let alone an exclusion for DIY e-bikes.
Often, in the US, your normal home-owners/renters policy has some level of umbrella liability coverage. Check with your plan to be sure.
The problem is (IMO) e-bikes that are more "motorcycle" than "bicycle". Which includes a massive number of the kits.
There currently is no national framework for classifying e-bikes. There's the 3 tier system that some industry groups use, but it directly conflicts with most state's moped/motorcycle regulations.
Safe bet for an e-bike is a "class 1" bike from a major brand. 20mph cap, no throttle, the most "bicycle" of the 3.
Class 2 bikes keep the 20mph cap, but add a throttle (don't need to pedal). This probably makes it a moped or small (50cc) scooter in some states.
Class 3 removes the throttle, but bumps the top speed to 28mph. Again, this speed probably makes it a moped or scooter (or possibly even a full motorcycle).
And then there's the e-bikes that are more motorcycle than bike. 30+mph, powerful engines, and the pedals are truly vestigial. Supe73 and Surron bikes fall in this category.
Huh, interesting that regulation exists (and since 2008?), as basically nobody follows that 20mph speed limit. Certainly the options from Specialized, Trek, etc usually provide assist to 28mph.
I’m ambivalent about throttles - much more concerned about the mass and speed of some e-bike when they inevitably get used on sidewalks, multi-use paths, etc.
Fair points, but that’s exactly my point—you’re assuming rather than knowing for sure. Medical insurance probably covers you, but some policies exclude high-risk activities or DIY mods, so it’s not always that simple.
And if you don’t think you have liability coverage, that’s exactly the risk. If your policy doesn’t mention bikes, it’s more likely not covered than automatically included. The more you do things outside the norm—like DIY e-bikes—the higher the chance standard policies don’t cover it.
Not saying you’re wrong, just that it’s worth checking so you don’t get caught off guard.
In my case (not US), home insurance cover bikes but not ebikes or escooters. For that you need a dedicated plan (<50€ per year). If I remember correctly, without that additional plan, I'm covered for my own injuries, but I'm on the hook for any injury I may cause in case of accident.
I _just_ went through this when my child had an ebike accident (AAA, southern california). You need specific ebike insurance. Auto/homeowners doesn't apply for different reasons.
Oddly, yes. A regular bike accident would have been covered. And it would have been covered on a moped as well even though she's too young to have a license.
Here it's easy to specify that a endpoint is a "stream", and then the code-generation tool gives all tools really to just keep serving the client with multiple responses. It looks deceptively simple. We already have setup auth, logging and metrics for gRPC, so I hope it just works off of that maybe with minor adjustments.
But I'm guessing you don't need the gRPC layer to use HTTP/2 Multiplexing?
At least in a browser context, HTTP/2 doesn't address server to client unsolicitied messages. So you'd still need a polling request open from the client.
HTTP/2 does specify a server push mechanism (PUSH_PROMISE), but afaik, browsers don't accept them and even if they did, (again afaik) there's no mechanism for a page to listen for them.
But if you control the client and the server, you could use it.
gRPC as specced to ride directly on top of HTTP/2 doesn't work from browsers, the sandboxed JS isn't allowed that level of control over the protocol. And often is too low level to implement as part of a pre-existing HTTP server, too. gRPC is a server-to-server protocol that is not part of the usual Web, but happens to repurpose HTTP/2.
Outside of gRPC, just HTTP POST cannot at this time replace websockets because the in-browser `fetch` API doesn't support streaming request body. For now, websockets is the only thing that can natively provide an ordered stream of messages from browser to server.
To understand how big this is, you have to look at the numbers:
The Naive Method: This is what most people learn in school. To multiply two 4x4 matrices, you need 64 multiplications.
The Human Record (1969): For over 50 years, the "gold standard" was Strassen’s algorithm, which used a clever trick to get it down to 49 multiplications.
The AI Discovery (2022): AlphaTensor beat the human record by finding a way to do it in just 47 steps.
The real "intelligence explosion" feedback loop happened even more recently with AlphaEvolve (2025). While the 2022 discovery only worked for specific "finite field" math (mostly used in cryptography), AlphaEvolve used Gemini to find a shortcut (48 steps) that works for the standard complex numbers AI actually uses for training.
Because matrix multiplication accounts for the vast majority of the work an AI does, Google used these AI-discovered shortcuts to optimize the kernels in Gemini itself.
It’s a literal cycle: the AI found a way to rewrite its own fundamental math to be more efficient, which then makes the next generation of AI faster and cheaper to build.
https://deepmind.google/blog/discovering-novel-algorithms-wi... https://www.reddit.com/r/singularity/comments/1knem3r/i_dont...
reply