The 'Exploring Image Models' section addresses that specifically, particularly the 'What Neurons Do Under Pressure' subsection.
'If it switched to superposition, we should see overlapping mixtures on top of each other, ... But that’s not what we see. The neurons stay interpretable and generalized, but coarser'
I've been waiting for symbolic music generators for ages. This is so nice for musicians. Being able to intelligently translate parts, generate additional parts, generate sheet music from audio, generate variations on pieces, making pieces easier or harder.
Spent some time playing on it just now on Linux. I couldn't easily get the automatic ingestion or gardening working, but everything else is working okay.
For my workflows I think I actually prefer wiki updates being manual and deliberate.
This isn't as easy as it sounds. Every ML model is struggling to balance between generalization and test performance.
Taking a good model like GLM5.2 and just fine tuning it on coding can decrease real world performance due to mechanics like catastrophic forgetting. There is also other interesting behaviors were training on a broad training set can improve coding performance because there is positive transfer.
There is 100% an effort to make solid coding focused models, but it is very hard to do that without including capabilities across a broad set of adjacent tasks.
Help me understand this viewpoint that AGI being possible in the near-ish future is a myth, I see it repeated quite a lot.
I've been in NLP since the LSTM days and it's hard for me to look at LLMs and not just think they are incredible. It's truly a different level of expressiveness. So much of capabilities research is pointing to LLMs effectively learning a world model.
RLVR is also proving really effective. It is hard for me to imagine a world in the future where LLMs aren't at human level performance across a wide variety of tasks.
I fully acknowledge that current LLM labs have a financial interest in people believing AGI is very near, but from what I'm reading in the literature and seeing myself experimenting with the SOTA models it doesn't seem totally unreasonable.
What evidence are you seeing that makes you confident that AGI in the soon-ish future is a complete myth?
I think you might be thinking of "AGI" as some sort of point in time, where something happens and everyone all at once has some technology. Not only is the progress towards AGI gradual, it's also very jagged in both capabilities and especially who has access to it. It's irrelevant whether a particular company, like Cisco, Pepsi, or Tesla, has some capability, when there exists a different research lab that is at the frontier, approaching AGI from some direction.
There doesn't really seem to be anything of substance in the actual executive order.
Section 1 doesn't say anything
Section 2 seems to boil down to: "improve cyber security and maybe use AI if we can find funding for it"
Section 3 proposes building a benchmark for evaluating cyber security performance of models that developers can choose to benchmark against. This seems like a good idea, I know Jack Clark has been a huge advocate for government's getting in with benchmarking.
Section 4 says to prioritize prosecuting cyber crimes. Not sure why they wouldn't already be prosecuted.
> Section 4 says to prioritize prosecuting cyber crimes. Not sure why they wouldn't already be prosecuted.
Not a whole lot of federal prosecutors. They're very selective about what gets pursued or not.
If they can't reliably build cases with a >90% success rate, it doesn't get prioritized. There's like <500 (federal) convictions per year on this whole area.
We hear about a few big famous ones in the news here, but most of it goes completely unenforced.
Not to mention quitting in droves because very many don't want to take these cases or otherwise to stand in court and explain why current admin is not bound by existing laws, court orders, the US constitution in general, or internationally recognized human rights etc.
If a federal prosecutor doesn't want to prosecute federal crimes, it's probably best for both themselves and their country if they find themselves a new job.
Out of curiosity, did you willfully choose to not understand the circumstances that prosecutors are being forced to carry hundreds of cases, too many to even read before they are in court, and then they are forced to stand in front of judges and face contempt while they are asked to explain why the government, who the prosecutor has no real control over, is violating yet another judicial order?
It isn’t just a matter of prosecutors picking and choosing…it’s underfunding, DOGE, and then those that are left are treated as adversarial the moment they complain about conditions or case loads. (Just like your comment does.)
It is only when judgement is rendered that it becomes a federal crime. Until then it is only alleged. And guess what: this administration is alleging a lot of things that fail.
Ah hahah yea. Not too much need about hackers being prosecuted going around. Lot more news about hackers breaching companies though. Closure rate of law enforcement & prosecutors vs hackers has gotta be way under 1% lol.
I’ve read lots of executive orders and it’s pretty standard. They don’t have much power. They are mostly just mandates and guidance for federal agencies, most of which is non binding, like a glorified mission statement. They just get sold as something bigger in the press.
Most voters don’t understand how the US government works, so EOs seem to be a way to pretend that the executive can pass laws. A way to make good on the campaign promises that require laws to be passed, which is usually all of them.
There is no actual regulation in EO 14319. It only covers federal government purchasing and vendor management. No one is required to change the "ideology" of an LLM, although they might not be able to sell it to the government.
That's not accurate. The EO explicitly lays out the implementation
> Sec. 4. Implementation. (a) Within 120 days of the date of this order, the Director of the Office of Management and Budget (OMB), in consultation with the Administrator for Federal Procurement Policy, the Administrator of General Services, and the Director of the Office of Science and Technology Policy, shall issue guidance to agencies to implement section 3 of this order.
A major LLM that did not submit to this would be labeled a "supply chain risk". It's unquestionable that every major LLM would go through this process
It even then goes on to say that existing contracts will be reviewed to ensure they are in compliance (reviewed by OMB)
> (b) Each agency head shall, to the maximum extent consistent with applicable law:
> (i) include in each Federal contract for an LLM entered into following the date of the OMB guidance issued under subsection (a) of this section terms requiring that the procured LLM comply with the Unbiased AI Principles and providing that decommissioning costs shall be charged to the vendor in the event of termination by the agency for the vendor’s noncompliance with the contract following a reasonable period to cure;
> (ii) to the extent practicable and consistent with contract terms, revise existing contracts for LLMs to include the terms specified in subsection (b)(i) of this section; and
> (iii) within 90 days of the OMB guidance issued under subsection (a) of this section, adopt procedures to ensure that LLMs procured by the agency comply with the Unbiased AI Principles.
You seem to be trying to win a technical argument while ignoring the practical implications of the order. But even technically I think you are wrong. Just because it's not passed by Congress doesn't mean it's not "regulation". OMB Memorandum M-26-04[0] is absolutely regulation and was created exactly because of this order.
Other agencies like the NIST (lookup NIST AI Risk Management Framework), the NAIAC (which was created in 2020) are the ones like in charge of:
> Agencies’ development of metrics, methods, and standards to test and measure AI, where such metrics, methods, and standards are for use by the general public or the Government as a whole, rather than to test AI for a particular agency application
Not technically but practically. The decrees are effectively considered law by the executive. Yes, you'll likely win in court later on, but you'll lose your job, get sent to prison, have your bank accounts and vehicle seized, etc., in the meantime.
Legality isn't really of much practical concern anymore. It's about what gets/can be enforced immediately.
What a weird comment. How many private business managers have ever been sent to prison for violating an EO? This particular EO doesn't even mention any criminal penalties.
Does the First Amendment actually let the US government dictate the types of speech you're allowed to put in your LLM? I mean, a US government that's bound by the Constitution, obviously.
While that issue hasn't been specifically tested in court yet, the current interpretation of the First Amendment probably wouldn't allow the US government to dictate the types of speech you're allowed to put in your LLM. But federal government purchasing decisions aren't generally bound by the First Amendment. In other words, government officials can generally refuse to purchase your LLM services if they don't like the speech it outputs. So there's no real constitutional concern with this EO.
I'm not claiming that this EO is sensible or enforceable, just that it's not prima facie unconstitutional.
Might be fair to say it’s setting the tone, though, that if you use “woke” (subjectively defined) ideology in any of your company’s marketing, documentation, or other communications you won’t be considered for government contracts. That’s a major blow for any company given the naked corruption and grift coming from the current admin.
It's not "setting the tone" it says that explicitly and even goes into detail into the implementation of how that is going to be enforced
> Implementation. (a) Within 120 days of the date of this order, the Director of the Office of Management and Budget (OMB), in consultation with the Administrator for Federal Procurement Policy, the Administrator of General Services, and the Director of the Office of Science and Technology Policy, shall issue guidance to agencies to implement section 3 of this order.
They even say they will review existing contracts
> (ii) to the extent practicable and consistent with contract terms, revise existing contracts for LLMs to include the terms specified in subsection (b)(i) of this section; and
> (iii) within 90 days of the OMB guidance issued under subsection (a) of this section, adopt procedures to ensure that LLMs procured by the agency comply with the Unbiased AI Principles.
The specific text reads like a favor to Elon Musk's xAI, since "Truth-seeking" is the buzzword Elon Musk frequently used to talk about Grok:
Sec. 3
Unbiased AI Principles.
It is the policy of the United States to promote the innovation and use of trustworthy AI. To advance that policy, agency heads shall, consistent with applicable law and in consideration
of guidance issued pursuant to section 4 of this order, procure only those LLMs developed in accordance with the following two principles (Unbiased AI Principles):
(a) Truth-seeking. LLMs shall be truthful in responding to user prompts seeking factual information or analysis. LLMs shall prioritize historical accuracy, scientific inquiry, and objectivity, and shall acknowledge uncertainty where reliable information is incomplete or contradictory.
(b) Ideological Neutrality. LLMs shall be neutral, nonpartisan tools that do not manipulate responses in favor of ideological dogmas such as DEI. Developers shall not intentionally encode partisan or ideological judgments into an LLM's outputs unless those judgments are prompted by or otherwise readily accessible to the end user.
> No one is required to change the "ideology" of an LLM, although they might not be able to sell it to the government.
It takes hundreds of millions of dollars to train a model. If a large customers, the government, says they won't buy it if it doesn't adhere to a set of standards, no one is "required" to change, but it's a pretty heavy hand on the lever. It's like having a job. No one is "required" to have a job, but having one gets you money, and having money is pretty important to modern life.
So what? No business has a legal right to sell their products to the government. The LLM vendors should probably find cheaper, more efficient ways to train their models rather than depending on government contracts. If your business lives or dies based on a single large customer then you don't have a viable business in the first place.
Are we looking at the same data? On that site I see that opus 4.7's and gpt 5.5's g scores are within each others confidence intervals, and both significantly ahead of the number 3 model.
Your comment makes it sound like they are miles apart, which the benchmark doesn't seem to support.
Edit:
I looked at the data more and the two models are only basically equal when looking at the mean of all the tests. Gpt 5.5 significantly outperforms opus 4.7 in coding, while opus 4.7 significantly outperforms in "decision making." I'm not seeing details on what decision making explicitly means.
Decision making refers to the environments where the LLM is called on every tick (like games with social communication), examples here: https://gertlabs.com/spectate.
Because GPT 5.5 just launched and those games take longer to accumulate data for, it just doesn't have enough samples yet. It will end up with a wider lead on Opus, I am sure. Coding evals always have large sample sizes on day 1. Good find, we should probably better adjust the weighting here for decision games with low match counts.
Right, I'm including my own observations in what the leaderboard is showing. Could be confirmation bias, but I use both Opus and GPT extensively and since GPT 5.4 I have noticed that Opus doesn't even begin to touch GPT's level of technical depth. I was hoping Opus 4.7 would close that gap, but unfortunately it doesn't even compare to GPT 5.4 in that sense.
I'm not being a hater, I love Opus for different reasons, but I can't rely on it for its technical ability.
If you take a chart of population density, and overlay the chart of suicide rate, you'll see an exceedingly strong correlation. It does not follow weather patterns. Utah has 3x the problem relative to California, for example.
Yeah if you go to CDC WISQARS you can do fatal injury reports filtered by intent (suicide) and aggregated by urban/non urban geography. These differences are not small, they vary by factors or orders of magnitude in every state. It's not the weather.
'If it switched to superposition, we should see overlapping mixtures on top of each other, ... But that’s not what we see. The neurons stay interpretable and generalized, but coarser'
reply