I'm rather convinced that the next major language-feature wave will be permissions for libraries. It's painfully clear that we're well past the point where it's needed.
I didn't think it'll make things perfect, not by a long shot. But it can make the exploits a lot harder to pull off.
The headline feature isn’t the 25 MB footprint alone. It’s that KittenTTS is Apache-2.0. That combo means you can embed a fully offline voice in Pi Zero-class hardware or even battery-powered toys without worrying about GPUs, cloud calls, or restrictive licenses. In one stroke it turns voice everywhere from a hardware/licensing problem into a packaging problem. Quality tweaks can come later; unlocking that deployment tier is the real game-changer.
You're arguing here as if there is an argument to be had. I'm trying to help you understand that there is no such argument. Your claims have been considered and rejected hundreds of times over the last 10 years. It'd behoove you to consult the search bar below and read some of those old discussions. This isn't a case where the community made a rash decision and just doesn't realize how problematic it is; you're talking about what is probably the single most deliberate community management decision HN has ever made. It's not going to change.
A moment later
Maybe it'll help if you get your head around the goal of HN. There are a bunch of things you personally seem to want to accomplish with HN. But HN itself has just one goal: to nurture curious conversation. Other sites have other, equally valid goals: propagating the news as quickly as possible, or relentless focus on a particular topic, or getting the best Q&A pairs to the top of Google's search results. HN, though: curious conversation. That's the whole ballgame.
Anything that drags on curious conversation is a non-starter on HN, even if it might make a bunch of things better for you.
We don't have to guess about whether making comment scores available is a drag on curious conversations; it manifestly was, for years. People have an innate, visceral reaction to comment scores they feel are unjust, and they talk about them, and those conversations choke the landscape like kudzu.
A lot of things about HN make more sense when you accept the premise of the site, and understand that HN will make most sacrifices it can come up with to optimize for that premise.
Baba is You is a great game part of a collection of 2D grid puzzle games.
(Shameless plug: I am one of the developers of Thinky.gg (https://thinky.gg), which is a thinky puzzle game site for a 'shortest path style' [Pathology] and a Sokoban variant [Sokoath] )
These games are typically NP Hard so the typical techniques that solvers have employed for Sokoban (or Pathology) have been brute forced with varying heuristics (like BFS, dead-lock detection, and Zobrist hashing). However, once levels get beyond a certain size with enough movable blocks you end up exhausting memory pretty quickly.
These types of games are still "AI Proof" so far in that LLMs are absolutely awful at solving these while humans are very good (so seems reasonable to consider for for ARC-AGI benchmarks). Whenever a new reasoning model gets released I typically try it on some basic Pathology levels (like 'One at a Time' https://pathology.thinky.gg/level/ybbun/one-at-a-time) and they fail miserably.
Simple level code for the above level (1 is a wall, 2 is a movable block, 4 is starting block, 3 is the exit):
000
020
023
041
Similar to OP, I've found Claude couldn’t manage rule dynamics, blocked paths, or game objectives well and spits out random results.
Apparently human language conveys information at around 39 bits/s. You could use a similar technique as that paper to determine the information rate of a speaker and then correct it to 39 bits/s by changing the speed of the video.
The key insight is that you can represent different features by vectors that aren't exactly perpendicular, just nearly perpendicular (for example between 85 and 95 degrees apart). If you tolerate such noise then the number of vectors you can fit grows exponentially relative to the number of dimensions.
12288 dimensions (GPT3 size) can fit more than 40 billion nearly perpendicular vectors.
No offense, but I wouldn't take Datalog 2.0's small attendance as an exemplar of Datalog's decline, even if I agree with that high-level point. Datalog 2.0 is a satellite workshop of LPNMR, a relatively-unknown European conference that was randomly held in Dallas. I myself attended Datalog 2.0 and also felt the event felt relatively sparse. I also had a paper (not my primary work, the first author is the real wizard of course :-) at the workshop. I myself saw relatively few folks in that space even attending that event--with the notable exception of some European folks (e.g., introducing the Nemo solver).
All of this is to say, I think Datalog 2.0's sparse attendance this year may be more indicative of the fact that it is a satellite workshop of an already-lesser-prestigious conference (itself not even the main event! That was ICLP!) rather than a lack of Datalog implementation excitement.
For what it's worth, none of what I'm saying is meant to rebut your high-level point that there is little novelty left in implementing raw Datalog engines. Of course I agree, the research space has moved far beyond that (arguably it did a while ago) and into more exotic problems involving things like streaming (HydroFlow), choice (Dusa), things that get closer to the general chase (e.g., Egglog's chase engine), etc. I don't think anyone disagrees that vanilla Datalog is boring, it's just that monotonic, chain-forward saturation (Horn clauses!) are a rich baseline with a well-understood engineering landscape (esp in the high-performance space) to build out more interesting theories (semirings, Z-sets, etc..).
I've -literally- been writing Swift, every day, seven days a week, 52.4 weeks a year, since June 2, 2014 (the day it was announced), yet, I still have huge gaps in my knowledge of the language.
I speak it without an accent, but not at Ph.D level.
As to home projects, that's pretty much all I do, these days, ever since I "retired"*, in 2017.
I'm quite good at what I do, and generally achieve every goal that I set, but, since I'm forced to work alone, the scope needs to be kept humble. I used to work as part of a worldwide team, doing some pretty interesting stuff, on a much larger scale.
But what's important to me, is that I do a good job on whatever I do. Everything I write, I ship, support, and document, even if it isn't that impressive. The bringing a project to completion, is a big part of the joy that I get from the work.
This is just a consequence of the fact that bfloat16 has a very high dynamic range which is not all used. People like hyperparameters that look like 0.01 not 10^10, even though there is the same fractional precision available at each exponent and if you multiplied everything - hyperparameters, initialized weights, training data, etc in a network by 10^6 things will still work more or less the same since the upper range is hardly used (with the possible exception of some small number of special functions).
Typical entropy of bfloat16 values seen in weights (and activations) are about 10-12 bits (only 65-75% or so of the value range is used in practice). Sign and mantissa bits tend to be incompressible noise.
This has been exploited several times before in the context of both classical HPC and AI, with lossless compression work from Martin Burtscher's lab (https://userweb.cs.txstate.edu/~burtscher/), fpzip from LLNL (https://computing.llnl.gov/projects/fpzip) and my library dietgpu from 2021 (https://github.com/facebookresearch/dietgpu) which we used to speed training on a large GPU cluster by about 10% wall clock time overall by losslessly compressing all data prior to send and decompressing upon receive (e.g., gradients, weights from backup, etc), which is still computing the same thing as it did before as it is lossless.
Also, rANS is more efficient and easier to implement in SIMD-like instruction sets than Huffman coding. It would reduce the performance latency/throughput penalties as well with DFloat11 (since we have to decompress before we do the arithmetic).
If anyone's interested, I made Colab notebooks with free GPUs for both GRPO (the algo DeepSeek used) to train a reasoning model from scratch, and also general finetuning, which the Berkeley team employed!
I'm serving AI models on Lambda Labs and after some trial and error I found having a single vllm server along with caddy, behind cloudflare dns, to work really well and really easy to set up
If anyone's interested and wants to hear more, I have a mix of 92/93 era Jungle [1]
Some rough mixes here and there (especially the first one) because it was live from a NYE event. But it suits the style of music, that era was so raw and fresh, the future was being invented right there! Very happy days :)
The era of startups isn't over. The era of low quality startups certainly is though, which is basically the point this article makes at the end among the general doomery. This in my opinion, is great news.
Being a startup can be conceived of as one possible path taken as an early phase in the business cycle, hard to see that changing. What is changing is now you have to be more realistic about justifying to others that you can be quickly scaled for a return. You need to have high natural growth underlined by great unit economics and the right plan to utilize capital to super charge that. Free money just meant wider bounds to take risks, which in a lot of cases turned into mediocre business. This is bad for everyone involved, VC down to consumer.
You now need to be more realistic about what is a lifestyle business or something that can be bootstrapped into a solid medium sized business through slow methodical growth instead of taking capital, flailing, reducing the product quality in a bid to survive, harm consumers and then dying anyway.
What I think this does harm unfortunately are moonshots, now you'll need to be even closer to some idealized SaaS with easily digestible metrics and plans to get funding. I imagine it's very very hard to separate a true moonshot that will succeed and revolutionize vs bunk and the risk profile has changed a lot, and quickly.
As a side note I also hope this reduces startup solely as a vehicle to acquisition and restores some novel public companies being created.
1. Datomic - While not open-source, it has an open-source version called Datomic Free, which is a distributed database designed to enable scalable, flexible, and intelligent data storage and queries. Datomic's query language is closely inspired by Datalog.
2. DataScript - An open-source in-memory database and query engine for Clojure, ClojureScript, and JavaScript that is heavily influenced by Datalog and Datomic.
3. Crux (now XTDB) - A bitemporal database with Datalog-inspired querying capabilities. It is designed for efficient querying of historical data and offers ACID transactions.
4. Racket's miniKanren - While not strictly a database, miniKanren is an open-source logic programming extension to the Racket language, which is inspired by Datalog and can be used to manipulate and query data in a manner similar to Prolog.
5. LogicBlox - An open-source platform that combines a database system, a Datalog-based modeling language, and application server facilities. It allows developers to build complex, data-intensive applications.
6. Soufflé - A Datalog-inspired language that is designed for static analysis problems. It can be viewed as a database query language with a focus on performance, allowing for parallel execution of queries.
7. Dedalus - A Datalog-like temporal logic language used to express complex distributed systems. It is primarily a research tool but has informed the design of other Datalog-inspired systems.
8. Flora-2 - An open-source object-oriented knowledge representation and reasoning system that integrates a variant of Datalog with objects and frames.
Top 3 are from the Clojure ecosystem. Additionnaly in this same space there is Datalevin & Datahike among many others
I didn't think it'll make things perfect, not by a long shot. But it can make the exploits a lot harder to pull off.