Thankfully the vast majority of the computing effort goes into pretraining. Censorship and vendor instilled bias applied on top especially in smaller models is relatively easy to undo. Give it a few weeks and I'm sure there will be tons of tweaked versions on huggingface with much of the brain damage removed.

You can just download it...

https://huggingface.co/Nikolay...

My opinion it's by far the best 3b model I've seen.

You tell me, the term is defined in the text of the bill. What don't I understand?

(3) FOREIGN ADVERSARY CONTROLLED APPLICATION.--The term "foreign adversary controlled application" means a website, desktop application, mobile application, or augmented or immersive technology application that is operated, directly or indirectly (including through a parent company, subsidiary, or affiliate), by-
 
(A) any of--
(i) ByteDance, Ltd.;
(ii) TikTok;
 
(iii) a subsidiary of or a successor to an entity identified in clause (i) or (ii) that is controlled by a foreign adversary; or
 
(iv) an entity owned or controlled, directly or indirectly, by an entity identified in clause (i), (ii), or (iii); or
 
(B) a covered company that--
 
(i) is controlled by a foreign adversary; and
(ii) that is determined by the President to present a significant threat to the national security of the United States following the issuance of--
 
(I) a public notice proposing such determination; and
(II) a public report to Congress, submitted not less than 30 days before such determination, describing the specific national security concern involved and containing a classified annex and a description of what assets would need to be divested to execute a qualified divestiture.

While I'm no lawyer I do have some experience with logical operations. If A is always true then it is not possible for 'A OR B' to ever evaluate to false. So no the president cannot in fact remove companies from the list.

The fact congress specified one specific company and has an entirely separate regime for adding any other companies means the company they added was singled out for special treatment not applicable to anyone else.

It would be one thing to create a TikTok bill and never mention TikTok but when you have materially different inclusion criteria for one named organization that isn't treating everyone the same.

Polonium is readily available:
https://amstat.com/products/an...

Bing Copilot is not relevant, it was not used.

The point isn't that model had access to the CVE it's that it was able to create an exploit using it.

With a model that predates the knowledge of CVE the expliot itself is unlikely to be contained within its training set therefore the work to derive an exploit would have had to be carried out by the model.

CVE descriptions contain information about a problem and are often intentionally vague. They normally would not reveal how to perform the exploit. That you have to investigate yourself by diffing old and new versions of source code or scrutinizing the code or probing system based on provided hints contained in the CVE description.

The text of the bill literally singles out ByteDance, Ltd and TikTok by name as foreign adversary controlled applications.

"FOREIGN ADVERSARY CONTROLLED APPLICATION.--The term "foreign adversary controlled application" means a website, desktop application, mobile application, or augmented or immersive technology application that is operated, directly or indirectly (including through a parent company, subsidiary, or affiliate), by--

(A) any of--
(i) ByteDance, Ltd.;
(ii) TikTok;
"

The only other way provided for in the bill to get this designation is POTUS determining there is a significant threat to the national security of the United States.

Calling out a company by name exposing it to a different standard than all others does not strike me as upholding equal protection. You can argue the constitutional protections don't apply to foreign corporations despite the US nexus... yet clearly TikTok is very much being treated differently.

The ultimate capabilities of a model is not merely a function of model size because these systems (especially /w agent based augmentation including long term storage and model directed tool use) are able to decompose problems, leverage the outcomes of previous computations and direct external processing.

LLMs have been demonstrated to be Turing complete.
https://arxiv.org/pdf/2301.045...

There is no such thing as a Turing machine in the real world as it requires an infinite tape which is not physically possible. You can only ever create a machine capable of acting on an external memory of infinite extent.

Or to put it in different terms try executing a Turing machine to do basic arithmetic in your head without a paper and pencil and let us how that goes.

Nobody is talking about AGI.

It is obvious from the description CVE descriptions were uploaded into the context as I said earlier. "From the paper I assume the advisories were uploaded into the models context" Context is roughly similar to a short term memory. Context doesn't change or augment the weights of the underlying model in any way. It is basically just part of the chat log / "prompt" transmitted to the model.

Again the API does not have the ability to search the web. The agent may well be doing that for all we know... they don't say... but the API itself is NOT the same thing as Bing search and does NOT have access to search the web.

This is like saying you know how processors work therefore you know how the software that runs on them works.

Several months ago I asked an instruction tuned deepseek model to write a program in a DSL it has never seen before. The only possible way it would have been able to produce a working program is by sufficiently understanding the language documentation uploaded into the models context and applying that understanding to produce a valid, properly formatted working program that did what I requested it to.

It's true that LLMs are limited and unreliable, they don't think like humans do yet they are very much able to apply concepts and knowledge to solve problems. In this case there isn't just an LLM but an agent framework with tooling demonstrated to substantively enhance the capabilities of LLMs. It can call on tools to gather additional data. If the LLM fucks up it can learn from its failure. If it goes down the wrong rabbit hole the agent can backtrack and try something else.

From the paper I assume the advisories were uploaded into the models context:

"When given the CVE description, GPT-4 is capable of exploiting 87% of these vulnerabilities compared to 0% for every other model we test "

"Fortunately, our GPT-4 agent requires the CVE description for high performance: without the description, GPT-4 can exploit only 7% of the vulnerabilities. "

They used an agent that calls GPT-4 via the API. API is only capable of querying the model and does not have web access.

"We use the ReAct agent framework as implemented in LangChain. For the OpenAI models,
we use the Assistants API."

It's a shame the paper does not seem to include useful information about the actions taken by this LLM driven agent. For all we know the agent hired a human to do the work for it or looked up the answers online. If one assumes no "cheating" then the results are impressive because the model would likely not have been trained on the answers.

"We further note that GPT-4 achieves an 82% success rate when only considering vulnerabilities after the knowledge cutoff date (9 out of 11 vulnerabilities)"

Fuses, circuit breakers are way too slow to be of any use during highest energy components of a nuclear EMP. Even MOVs are too slow. Would need something like TVS diodes to respond quickly enough.

I think the best source of data is still the EMP commission report because they actually do real world testing of control systems, computers, network cables, vehicles..etc rather than just calculations and conjecturbation.

https://apps.dtic.mil/sti/pdfs...

Apart from the y'all-watch-this factor, security-wise it seems it'd be more useful to have a completely different, randomised build each time so attackers can't target a monoculture binary image.

In terms of "you can use it to verify source to binary equivalence", you're already relying entirely on trusting the developers to not do anything malicious, so what advantage is there to a reproducible build vs. downloading a signed binary? And for it to work you need signed source code and a signed attestation that when built the code has hash value X, which is just a really clunky way to do a signed binary.

What attack is this preventing that manifests and signed binaries as used by every package manager ever for years don't?

If you read mussolini's manifesto fascism is basically stateism / totalitarianism with aggressive expansionism (usually but not exclusively militarily) baked in. The gist is that everything in society is organized around and in service to the state/leader which takes precedence over all else. Basically the fever dream of a petty tyrant who thinks the world revolves around him.