LeCun's technical approach with AMI will likely be based on JEPA, which is also a very different approach than most US-based or Chinese AI labs are taking.
If you're looking to learn about JEPA, LeCun's vision document "A Path Towards Autonomous Machine Intelligence" is long but sketches out a very comprehensive vision of AI research:
https://openreview.net/pdf?id=BZ5a1r-kVsf
Training JEPA models within reach, even for startups. For example, we're a 3-person startup who trained a health timeseries JEPA. There are JEPA models for computer vision and (even) for LLMs.
You don't need a $1B seed round to do interesting things here. We need more interesting, orthogonal ideas in AI. So I think it's good we're going to have a heavyweight lab in Europe alongside the US and China.
Rest in peace, he hasn't seen the industry change.
"A consequence of this principle is that every occurrence of every subscript of every subscripted variable was on every occasion checked at run time against both the upper and the lower declared bounds of the array. Many years later we asked our customers whether they wished us to provide an option to switch off these checks in the interests of efficiency on production runs. Unanimously, they urged us not to they already knew how frequently subscript errors occur on production runs where failure to detect them could be disastrous. I note with fear and horror that even in 1980 language designers and users have not learned this lesson. In any respectable branch of engineering, failure to observe such elementary precautions would have long been against the law."
-- C.A.R Hoare's "The 1980 ACM Turing Award Lecture"
I think a more likely reason, that for some reason, a lot of people don't want to talk about, is that these "Global Elite" aren't really that smart, creative, or articulate. That they've gotten to where they are despite, not because of their communication skills. They're not being "typical unconventional / quirky entrepreneurs." They're simply C students who knew the right people.
Hi, author of this devlog here! Not to dismiss concerns about breaking language changes, but there seems to be a bit of a misconception here that this compiler change was highly breaking and will require significant effort from Zig users to update for. Perhaps I unintentionally gave that impression in the devlog or the PR writeup, apologies if so---but it's not the case! Although there were breaking changes in this patch, they were quite minor: most users are unlikely to hit them, and if they do then they're straightforward to deal with.
For a concrete example, while testing this branch, I tried building ZLS (https://github.com/zigtools/zls/). To do that, the only change I had to make was changing `.{}` to `.empty` in a couple of its dependencies (i.e. not even in ZLS itself!). This was needed because I removed some default values from `std.ArrayList` (so the change was in standard library code rather than the language). Those default values had actually already been deprecated (with intent to remove) for around a year, so this wasn't exactly a new change either.
As another example, Andrew has updated Awebo (https://codeberg.org/awebo-chat/awebo), a text and voice chat application, to the new version of Zig. Across Awebo's entire dependency tree (which includes various packages for graphics, audio, and probably some other stuff), the full set of necessary changes was:
* Same as above, change `.{}` to `.empty` in a few places, due to removal of deprecated defaults
* Add one extra `comptime` annotation to logic which was constructing an array at comptime
* Append `orelse @alignOf(T)` onto an expression to deal with a newly-possible `null` case
These are all trivial fixes which Zig developers would be able to do pretty much on autopilot upon seeing the compile errors.
So, while there were a handful of small breaking changes, they don't seem to me like a particularly big deal (for a language where some level of breakage is still allowed). The main thing this PR achieved was instead a combination of bugfixes, and enhancements to existing features (particularly incremental compilation).
The title is misleading — there's no trained 100B model, just an inference framework that claims to handle one. But the engineering is worth paying attention to.
I run quantized 70B models locally (M2 Max 96GB, llama.cpp + LiteLLM), and memory bandwidth is always the bottleneck. The 1.58-bit approach is interesting because ternary weights turn matmuls into additions — a fundamentally different compute profile on commodity CPUs. If 5-7 tok/s on a single CPU for 100B-class models is reproducible, that's a real milestone for on-device inference.
Framework is ready. Now we need someone to actually train the model.
Geohot is the epitome of someone who thinks because they're exceptionally intelligent and competent in a niche area, they're in a position to confidently explain how the world "really" works, without having to put any effort into actually researching areas outside of their niche.
His blog posts and general opinions voiced in his streams in any other field than what he's working in are so incredibly stupid and put forward with so much misguided confidence that they make me cringe in pain.
I feel like that sort of verification is just inherently flawed and easy to bypass. I mean as easy as just telling your agent "hey go publish this on moltbook".
My pet theory is Meta got acquihire FOMO after seeing OpenAI acquire Openclaw/Peter Steinberger.
McKinsey has a weird structure where there are too many cooks in the kitchen.
Everybody there is reviewed on client impact, meaning it ends up being an everybody-for-themselves situation.
So as a developer you have little guidance (in fact, you're still being reviewed on client impact, even if you have 0 client exposure).
Then a (Senior) Partner comes in with this idea (that will get them a good review), and you jump on that. After all, it's all you can do to get a good review.
You work on it, and then the (Senior) Partner moves on. But it's not done. It's enough for the review, but continuing to work on it doesn't bring you anything, in fact, it will actually pull you down, as finishing the project doesn't give immediate client results.
So what does this mean? Most products of McKinsey are a grab-bag of raw ideas of leadership, implemented as a one-off, without a cohesive vision or even a long-term vision at all. It's all about the review cycle.
McKinsey is trying to do software like they do their other engagements. It doesn't work. You can't just do something for 6 months and then let it go. Software rots.
The fact that they laid off a good amount of (very good) software engineers in 2024 is a reflection on how they see software development.
And McKinsey's people, who go to other companies, take those ideas with them. Result: The UI of your project changes all the time, because everybody is looking at the short-term impact they have that gets them a good review, not what is best for the project in the long term.
Also relevant: The DOGE team set up a Starlink satellite at the White House [1].
DOGE staff installed the terminal on the Eisenhower Executive Office Building roof in February 2025 without notifying White House communications or cybersecurity teams, ignoring their prior warnings [2]. The resulting "Starlink Guest" Wi-Fi used only a password—no usernames or two-factor authentication—unlike standard networks requiring full VPN tunneling and device logging.
This allowed devices to evade monitoring, transmit untracked data outside secure channels, and potentially enable leaks or hacks, as noted by former officials and experts like ex-NSA hacker Jake Williams. A confrontation ensued with Secret Service when DOGE accessed the roof unannounced [3].
I come from the Minecraft modding/server community. There is interesting fact that I like to tell people about the sheer size of Roblox compared to other communities like Minecraft.
The largest Minecraft server in the world is Hypixel at around ~30K concurrent players. Most other servers are very far behind.
There is one Roblox game that looks and plays like Minecraft and copied one single gamemode (Bedwars) common in servers like Hypixel. It had 60K+ concurrent players last time I checked late last year.
There are almost definitely more people playing BedWars on Roblox than there are playing it on Minecraft at this very moment.
I've been wearing an Apple Watch for close to 10 years. I've tracked my weight as well along those years but nothing crazy like OP. The Apple watch tracked plenty.
I had some strange symptoms and two doctors insisted I had a weak heart and potential heart failure. This was shocking! Turns out I do have a really "weak" rhythm, but heart failure is when your heart is progressively getting worse in it's pumping. I don't even remember which metric he looked at in my Apple health - but basically my heart has always been this way. A doctor looking at a single data point might think I have abnormally low blood pressure/heart rate, but if I've had this for 10 years with no change, the medical assessment is very different - it means nothing. Sometimes boring data is exactly what you need. For this reason, I will probably always wear an Apple watch (or equivalent) moving forward.
Data can feel useless for 10 years until one day it becomes critical. The benefit is spiky and uneven.
Well, If the staff got stiffed on the fee "many times", and the parents were allowed to bring their kid back.. the place didn't charge $1 per minute late. They just bluffed and got called on it.
(apologies for the immediate edit, changed my wording)
The trick is, with the setup I mentioned, you change the rewards.
The concept is:
Red Team (Test Writers), write tests without seeing implementation. They define what the code should do based on specs/requirements only. Rewarded by test failures. A new test that passes immediately is suspicious as it means either the implementation already covers it (diminishing returns) or the test is tautological. Red's ideal outcome is a well-named test that fails, because that represents a gap between spec and implementation that didn't previously have a tripwire. Their proxy metric is "number of meaningful new failures introduced" and the barrier prevents them from writing tests pre-adapted to pass.
Green Team (Implementers), write implementation to pass tests without seeing the test code directly. They only see test results (pass/fail) and the spec. Rewarded by turning red tests green. Straightforward, but the barrier makes the reward structure honest. Without it, Green could satisfy the reward trivially by reading assertions and hard-coding. With it, Green has to actually close the gap between spec intent and code behavior, using error messages as noisy gradient signal rather than exact targets. Their reward is "tests that were failing now pass," and the only reliable strategy to get there is faithful implementation.
Refactor Team, improve code quality without changing behavior. They can see implementation but are constrained by tests passing. Rewarded by nothing changing (pretty unusual in this regard). Reward is that all tests stay green while code quality metrics improve. They're optimizing a secondary objective (readability, simplicity, modularity, etc.) under a hard constraint (behavioral equivalence). The spec barrier ensures they can't redefine "improvement" to include feature work. If you have any code quality tools, it makes sense to give the necessary skills to use them to this team.
It's worth being honest about the limits. The spec itself is a shared artifact visible to both Red and Green, so if the spec is vague, both agents might converge on the same wrong interpretation, and the tests will pass for the wrong reason. The Coordinator (your main claude/codex/whatever instance) mitigates this by watching for suspiciously easy green passes (just tell it) and probing the spec for ambiguity, but it's not a complete defense.
A simple back of the envelope calculation shows that Felix causes between 70 and 110 tonnes of CO2 emissions per year just from flying.
Paris accord says 1.5t per person per year, from all activities, Felix's flying alonre is ~10-15x current European yearly per person emissions and ~50-75x those compatible with +1.5C.
> And this is the best country in the world, with the best system of government, because private citizens can voice their disagreement with such actions, including by refusal to participate.
On the off chance other Americans were unaware of this: Other countries are democracies too (and many are better functioning)
FYI, Wiz investor and current Wiz board member Gili Raanan, head of Israeli VC Cyberstarts, has been (credibly) accused of paying bribes to major CISOs for buying software from their portfolio companies like Wiz.
My only caution is that good writers and LLMs look very similar, because LLMs were trained on a corpus of good writers. Good writers use semicolons and em-dashes. Sometimes we used bulleted lists or Oxford commas.
So we should make sure to follow that other HN rule, and assume the person on the other end is a good faith actor, and be cautious about accusing someone of using AI.
(I've been accused multiple times of being an AI after writing long well written comments 100% by hand)
So much of this started with the rise of the peer-review journal cartel, beginning with Pergamon Press in 1951 (coincidentally founded by Ghislaine Maxwell's father). "Peer review" didn't exist before then, science papers and discussion was published openly, and scientists focused on quality not quantity.
Then the OS is full of ads and pre-installed garbage “gaming-optimization-tool” or driver tools taking up 99% of a single core while being riddled with security holes.
But inevitably, some chucklehead comes along "wut? I can get <proceeds to type spec sheet> for half that! Have fun paying the apple tax, lol." Someone posted that on Ars yesterday, with a random Amazon link from Naikan, your name for quality computing. Or rather, "Naikan, your name for a quality trackpad, screen, and high-quality ABS case! Be sure to check out the $12,000 of 'bonus' software add-ons, no extra charge!". It's amazing someone can post that without the slightest hint of self-awareness.
> You cannot buy an x86 PC laptop in the $600–700 price range that competes with the MacBook Neo on any metric — performance, display quality, audio quality, or build quality. And certainly not software quality.
I would argue the opposite: while Apple hardware is generally excellent, it is the software that leaves to be desired. Apple has also been consistently pushing the industry in a dangerous direction (walled gardens with app stores, excessive power over developers and users). MacOS is also very behind Linux these days in terms of app compatibility (especially games).
I won't be buying a Neo before a compatible Linux distro is confirmed. If the stock OS can't be replaced for one reason or another, it's dead on arrival as far as I am concerned.
But you didn’t spend hundreds of hours on it, so when it did happen to be useful it seemed like an outsized benefit.
I would wager that for most people, most data about themselves will be useless and not worth collecting.
Of course you can’t know what data will be useless or not, so unless the cost of collecting it is minimal or nil (wearing a smart watch, writing down your weight each day/week), it’s probably not worth it.
Spending hundreds of hours to build a solution to capture all data about yourself to find interesting patterns has a huge assumption baked into it: that there are interesting patterns to find.
Can any of the administration's defenders explain to me how this is actually a good thing and not the exact thing people were warning about a year ago?
If you're looking to learn about JEPA, LeCun's vision document "A Path Towards Autonomous Machine Intelligence" is long but sketches out a very comprehensive vision of AI research: https://openreview.net/pdf?id=BZ5a1r-kVsf
Training JEPA models within reach, even for startups. For example, we're a 3-person startup who trained a health timeseries JEPA. There are JEPA models for computer vision and (even) for LLMs.
You don't need a $1B seed round to do interesting things here. We need more interesting, orthogonal ideas in AI. So I think it's good we're going to have a heavyweight lab in Europe alongside the US and China.