HN user
If it's true that in under 15 days of access significant improvements were realized in K3, then the moat of closed-weight models is far smaller than previously thought.
Doesn't bode well for the valuations of these labs.
And corn which was used to make energy (as ethanol) at a fraction of the land efficiency of solar.
https://www.anthropocenemagazine.org/2025/04/new-study-compa...
Assuming all 64 subagents were running for a full hour (the tweet states just under an hour):
Throughput Output tokens Output cost
---------------------------- ------------- -----------
40 tok/s (5.5 low) ~9.2M ~$275
55 tok/s (5.5 base) ~12.7M ~$380
70 tok/s (5.5 high) ~16.1M ~$485
750 tok/s (Sol Fast, $75/M) ~172.8M ~$13,000
Edit: better tok/s estimate buckets based on GPT 5.5 actual speeds since I couldn't find real benchmarks on 5.6 published anywhere. Also account for Sol Fast pricing.
Let's use an example of a GW AI deployment.
At $0.07/kWh, that costs $70,000 every hour in just electricity. $1.7 million /day. $613 million /year.
I had claude estimate the GPU cost of such a deployment:
To get racks per GW: a full NVL72 rack draws roughly 130-132 kW under full load. If a 1 GW facility runs ~715 MW of IT power (after a ~1.4 PUE for cooling), that's on the order of 4,000–4,500 racks. At $3.4M of compute hardware each, the GPU-system cost lands around $14–15 billion.
15 billion / 613 million / year = ~24.5 years til electricity costs catch up to the GPUs. Obviously electricity isn't 100% of OpEx, but I'd expect it to be the majority for AI deployments.
Regardless, if you can cut the $613 million/yr in half that's still massive savings.
All these points are valid, and OpenAI did a great job identifying potential risks, especially misuse and biases, at an early stage.
Many of the OpenAI employees who were focused on these risks in GPT-2 later founded Anthropic, notably Dario [1]. Since the beginning and continuing through today Anthropic describes itself as an "AI safety and research company" [2]
I'm not sure if the OpenAI of today has the same focus on safety, or if they do the minimum to not look irresponsible given Anthropic's effort.
Are teaching and software engineering even job categories that overlap enough that they should compete for the same pool of visas?
It seems to me things would be better if they were classified as different visa categories.
There was a change that took place in the late 70s it looks like
Unfortunately in the current market 32GB of ddr5 seems to run about $400 as 2x16gb DIMMS, and even more for 1x32GB DIMM (higher density chips are more expensive). So $600 really isn't much over market price, especially considering strix halo uses 8000MHz ram instead of the typical 6000 found in consumer dimms.
The premise is correct, but this particular source is definitely an ad for a blood testing program.
The [Vitamin D Receptor] VDR is part of the nuclear receptor superfamily of steroid hormone receptors, which are hormone-dependent regulators of gene expression. These receptors are expressed in cells across most organs. [1]
I started using similar approaches in the sonnet 3.5 era and found them incredibly useful at the time. The frontier lab models have gotten significantly better about their guesses over time, but I still sometimes turn to the technique if my own ideation is only about 80% of the way there, as the LLM's questioning can help me identify the blind spots that need more consideration.
No simulation is perfect, so ideally you have a feedback look constantly looking at new real-world data as it comes in and finding where the simulation has errors, and updating the simulation to improve the correlation between the simulation and the real world over time.
My guess is they did have flooded street sims but the correlation was much lower than expected, or the details of the situation being simulated (lighting, building locations, how dirty the water is, ...) were sufficiently different from the situation that was encountered that the sim based training didn't generalize to the new context.
I like this concept of adding additional context to the dependency graph of imports.
Any plans for tooling that would enable rules like "no code above tear <N> allowed in the production build"?
Intel employs over 9000 people in Israel [1].
There's significant chip development and manufacturing talent there, so it's not really all that surprising if Apple has a team there for those purposes.
Also Apple bought out Intel's modem division in 2019 [2], so chances are that group was already Israel based under Intel.
1. https://www.intel.com/content/www/us/en/corporate-responsibi... 2. https://www.apple.com/newsroom/2019/07/apple-to-acquire-the-...
Age gating the VPN age gates (pseudo-anonymous) access to 100% of the content on the internet. Regardless of whether or not you agree with it, age gating only the porn subset of the internet is a much smaller restriction.
those that do not have an "unavoidable expiration date"."
Try avoiding the heat death of the universe /s
A more accurate phrasing is: It's significantly less likely that one learns the portion of the work they offload to an LLM.
A random anecdote is that most of the people I know who went very far in theoretical math are relatively poor at basic mental arithmetic, because they always think in the abstract and offload addition and multiplication to the calculator. It doesn't mean they can't do it, they just aren't as practiced or as fast at it.
Amazon has far more roles than just software. PMs, FC area managers, managers - if your job involves writing anything you're expected to be using AI in some capacity.
Even if the ultimate measure is dollars most employers will attempt to predict which metrics of employment best correllate with dollars so they can predict how many people to hire
Any claim about "productivity is due to X" that doesn't define a timescale is either flawed or misleading. In fact all measures of anything need to be done across some meaningfully defined time scale to have any relevance.
github has a huge warning saying to never use pull_request_target to run user code
This is an area where documentation is necessary but not sufficient. Github needs to add some form of automated screening mechanism to either prevent this usage, or at the very least quickly flag usages that might be dangerous.
If your intestinal lining is healthy, there's probably not much to worry about.
If it hurts while just sitting in your gut that's probably a red flag that something about your digestive system is off.
That's still coordinated, but by the publicizing of the key
What process or mechanism would you prefer to use instead of coordinated disclosure?
I believe this premise that the cost of identification of vulnerabilities via diffs is going down over time begs the question "what do our processes need to look like if simply making the patch public is the disclosure?"
Current coordinated disclosure practices have a dependency on patching and disclosure being separate, but the gap between them seems to be asymptomatically approaching zero.
I believe their point was that:
"How likely is this diff a patch for an existing vulnerability?"
Seems to be an easier question to answer than
"Are there any new vulnerabilities introduced by this diff?"
In other words identifying that a patch is for a vulnerability is typically easier than finding the vulnerability in the first place.
If the parameter names mostly standardize across tools because the models learn to predict those names, then humans will also learn to predict those flag names so this actually has the potential to make tools more human friendly and easier to learn.
[O]rganizations which design systems (in the broad sense used here) are constrained to produce designs which are copies of the communication structures of these organizations.
— Melvin E. Conway 1967
In pretty much every case where I've previously thought "I wish we had a tool for this but I can't get the time funded to make it" I now just get ai to work on the tool in the background and check in on it whenever I have a few minutes of deadtime before / after meetings.
The benefits of the time savings of having progressily better tooling over time add up quickly.
The general concept of a hook with delayed payoff is far from new, and generally one of the better ways at keeping attention.
It's also exactly the Mr beast playbook, and got him to the largest channel on YouTube.
Any system attempting to capture human attention will use these techniques, nothing LLM-specific here at all.
As a 16 year wearer of mostly barefoot shoes, "barefoot" for me is about comfort in general day to day activity. It isn't a specialized tool and certainly isn't the obvious choice for extreme environments.
If I'm going bouldering I absolutely cram my toes into a tiny rock climbing shoe, because it allows me to stand on ledges I couldn't without the extra support from the shoe.
That being said, if barefoot generally feels good to you and you're not chasing the pinacle of performance it's probably a perfectly fine choice for your recreational runs.