ClusterMAX™ currently has approximately 90% coverage of the entire GPU market by GPU volume

We will re-evaluate and update our GPU Cloud ClusterMAX™ Tier list every 3-6 months

we view being on the “AMD Alliance Instinct Cloud Partners” list as not a good predictor of tiering well in ClusterMAX™.

there is only one GPU cloud, CoreWeave, that provides services at this tier

No part of SemiAnalysis’s compensation by our clients was, is, or will be directly or indirectly related to the specific tiering, ratings or comments expressed.

Models are typically rewarded solely for correct outcomes, not penalized for incorrect reasoning, enabling them to achieve accuracy through flawed logic.

In the Claude 4 release, Anthropic significantly reduced reward hacking by improving environments, clarifying reward signals, and implementing proactive monitoring.

Claude 3.7 Sonnet exhibited reward hacking by altering test cases rather than improving its code to pass original tests.

Reliable, scalable, easy to implement environments will be in extreme demand and we expect this to be a growing area for startups to operate in.

Solving reward hacking is of top importance to all of the labs and will draw on many ideas from the safety-oriented teams.

There is an entire security infrastructure that needs to underpin this as well, so the model is protected from external penetration or from trying to escape the environment.

The Trump administration needs to solve this failure from the Biden administration immediately

DeepSeek has ambitions to release a multimodal model in V4, but scarce compute is slowing progress.

The argument Blackwell needs to be sold into China is a false narrative

805k this year, 653k of those being 910C

Assuming no smuggling, China will be able to make less Ascends next year, not more.

Make your graders resistant to bypasses or hacks. The agent shouldn’t be able to easily “cheat” the eval.

Each trial should be “isolated” by starting from a clean environment.

No model we tested could complete it until it was given a compute budget of at least 30M tokens

as leading providers are now doing

every model plateaued within its usual budget

the fitted frontier trend is ~60% steeper when horizons are estimated at 50M tokens rather than 2.5M tokens per task

I’ve decided that now is the right time for me to hand over my day-to-day operational responsibilities at GDM

are super focused on the areas where we need to improve

Flash is in high demand, our Cyber model is live, and Gemma models have surpassed 900M+ downloads

The Gemini models are in good hands with Koray and the leads, as they have been for a while

Jeff and Google Senior Fellow Sanjay Ghemawat are launching an independent public benefit corporation to accelerate discoveries in ML, science, and engineering.

We’ve also helped the US Centre for AI Standards and Innovation (CAISI), Model Evaluation and Threat Research (METR) and Apollo Research adopt the open-source sandbox providers for their own agentic evaluations

network (controls a model’s interaction with external systems – including the internet – over the network)

Every successful breakout exploited a previously disclosed vulnerability

one model found and exploited an unintended escape path, illustrating how easily gaps creep into container setups even when designed by security-conscious teams

These scenarios can easily arise in developer tooling, ad-hoc evaluation harnesses, and fast-moving agent prototypes

there did not appear to be anything special about the “American interests” goal we used in our main experiments

the models may simply assume they are in a real setting, not considering whether they are in an evaluation

In the cases where it stated the scenario was real, it blackmailed 55.1% of the time, compared to only 6.5% of the time when it stated it was in an evaluation

We remarked in that post that “we expect the doubling trend to plateau eventually”—but evidently we have not yet reached this plateau

Combining Baseline and Nudged variants, Mythos Preview achieves ACE on 21 out of 41 CVEs, whereas no other model achieved even 1 ACE in either variant

with contributions from security researchers at Anthropic, OpenAI, and Google

Whereas OpenAI’s models exploited a novel vulnerability to escape isolation, the Claude models evaluated here accessed the internet via an open path

the safeguards deployed on our generally available models would have blocked the behaviors identified

After reviewing 141,006 evaluation runs where Claude could have obtained internet access

consistent with a blameless postmortem culture, we’re approaching the fixes as if the responsibility were ours alone

Due to a misunderstanding between us and our evaluation partner, this was not the case, and internet access was available

within the next week, we will release a lightly redacted transcript in which Claude built a malicious PyPI package

Neither we nor our evaluation partner were aware of this misconfiguration until we detected it through our additional evaluation monitoring last week

our latest model stopped once it recognized it was on the internet

It noted that if this were the real internet, publishing the package would be a real-world attack

none stopped the attack on that basis

the calendar date of 2026 on the systems proved, according to Claude, that the environment was staged

I don't usually use these terms. It is evil. It's an evil world

I think it runs the world right now or it's on the brink of running the world because I think you see these kinds of characters elsewhere. This is not a simply an American phenomena.

On Sunday, when Willi Castro homered in a blowout Twins loss, Provus paid homage to Uecker’s home run call:“Castro in the air to right. Hey, get up, get up! Get outta here! Gone!” Provus said on the Twins TV broadcast.

freedom to run the program as you wish, for whatever purpose.

in the past these collapses of civilizations were local and people could migrate a little further on and rebuild but the chances of of that are gone now i mean we have to we have to uh to to 01:03:18 uh get right with what we have because it's all we have you know we we all all those bets we placed when our ancestors invented civilization they all rest on one high stakes throw which is 01:03:32 now

a run of T S is a computation that starts froms0.

Besides that ffscreencast can act as an ffmpeg command generator. Every available option can also just show the corresponding ffmpeg command instead of executing it. Non-ffmpeg commands, such as how the camera resolution is pulled and others can also be shown instead of being executed.

I resolved this by changing the imports from @remix-run/node to @remix-run/cloudflare

This is an effective, dynamic stand-in for a fixed secret.

opts = method(__method__).parameters.map(&:last).map { |p| [p, eval(p.to_s)] }.to_h SomeOtherObject.some_other_method(opts)

that's right, we don't want to do params = { ... } because then we're hardcoding the implementation and it becomes very coupled. The benefit of doing it like in my examples is that you can change the method signature and still automatically capture all keyword parameters.

Prefer alias when aliasing methods in lexical class scope as the resolution of self in this context is also lexical, and it communicates clearly to the user that the indirection of your alias will not be altered at runtime or by any subclass unless made explicit.

He's also the co-founder of the hyperlocal community site outside.in.

The type-fest package contains only types, meaning they are only used at compile-time and nothing is ever compiled into actual JavaScript code. This package contains functions that are compiled into JavaScript code and used at runtime.

Introduced in the perfectly named “Typescript and validations at runtime boundaries” article @lorefnon, io-ts is an active library that aim to solve the same problem as Spicery:TypeScript compatible runtime type system for IO decoding/encoding

It means that when having a type guard:TypeScript and JavaScript runtime are tied to the same behaviour.

Inside the if statement, TypeScript will assume that amount cannot be anything else than a string, which is true also at the runtime thanks to typeof JavaScript operator.

SQL is all about nested subqueries. It's hard to escape without creating views, but who has time to lookup that syntax and get their DBA's permission to run the DDL?!?

This marks the transition from load time (and dynamic link time, if present) to run time.

Since we're not passing any inputs to ListAccounts, it makes sense to use .run! instead of .run. If it failed, that would mean we probably messed up writing the interaction.

As a workaround, I guess I'll have to disable my strict CSP in development, but I'd prefer to keep it strict in development as well so that I ran into any CSP issues sooner...

In order to invoke, or run

I'd say it's better overall because it replaces paths at compile-time. It means there is no runtime dependency or any performance overhead.

Loaders use a mapping configuration to map module names to files at run-time, see RequireJs documentation and SystemJS documentation.

Note that you can also use dynamic partials, that conditionally select the partial to render based on the value in the JSON.

we update the validation schema on the fly (we had a similar case with a validation that needs to be included whenever some fetch operation was completed)

What would happen if we get the list from a rest api when the form component is mounted? That's an interesting topic, you can add a rule once a component has been mounted and update the associated validation schema.

Ok, I have seen that there are lot of built-in and third party validations, but sooner or later I will face a validation rule not covered by this buffet. Can I build a custom one? Of course you can!

Doing so also means adding empty import statements to guarantee correct order of evaluation of modules (in ES modules, evaluation order is determined statically by the order of import declarations, whereas in CommonJS – and environments that simulate CommonJS by shipping a module loader, i.e. Browserify and Webpack – evaluation order is determined at runtime by the order in which require statements are encountered).

This PR replaces nested require statements in files in vendor/postcss with import declarations, so that they can be properly resolved at build time.

Svelte will not offer a generic way to support style customizing via contextual class overrides (as we'd do it in plain HTML). Instead we'll invent something new that is entirely different. If a child component is provided and does not anticipate some contextual usage scenario (style wise) you'd need to copy it or hack around that via :global hacks.

Explicit interfaces are preferable, even if it places greater demand on library authors to design both their components and their style interfaces with these things in mind.

The point of the feature is to not rely on the third-party author of the child component to add a prop for every action under the sun. Rather, they could just mark a recipient for actions on the component (assuming there is a viable target element), and then consumers of the library could extend the component using whatever actions they desire.

Your LazyLoad image is now inextensible. What if you want to add a class? Perhaps the author of LazyLoad thought of that and sets className onto the <img>. But will the author consider everything? Perhaps if we get {...state} attributes.

You'll have to create a new component that brings in the functionality of both. TooltipButton, TooltipLink, Link, and TooltipRoutedLink. We're starting to get a lot of components to handle a bit of added functionality.

For the tooltip example, if you had a whole bunch of tooltips on different elements, it would be annoying to have different event listeners and "should it be shown" variables for each one.

Perhaps at that point we're better off settling on a way to pass components through as parameters? <!-- App.html --> <Outer contents={Inner}/> <!-- Outer.html --> <div> <div>Something</div> <[contents] foo='bar'/> </div>

I would hope for it to come with React-like behavior where I could pass in a string (like div or a) and have it show up as a normal div/a element when the child component used it.

Use case: Wrapper components that need to render an element (e.g. because they attach event listeners). You'd probably use a <div> there by default but there may be places where this is not desirable for semantic reasons (e.g. in lists).

const components = { Label, Tree, Menu };

The lack of spread continues to be a big pain for me, adding lots of difficult-to-maintain cruft in my components. Having to maintain a list of all possible attributes that I might ever need to pass through a component is causing me a lot of friction in my most composable components.

Yes, traditional scenario load tests are naturally in the risk zone of being axed in the name of this-step-is-taking-too-long as load tests need time to ramp-up and execute the user journeys with the simulated traffic to gain enough measurements that can be acted on. This is why we don’t recommend load tests to be run on every commit for scenario type load tests, but rather in the frequency range of “daily” for performance regression type tests. When merging code into a release branch or as a nightly build perhaps, so that you can have your snazzy load test report with your morning coffee before you’ve settled into your zone!

An ssh public key in a ~/.ssh/authorized_keys file can have a command="" option which forces a particular command to be executed when the key is used to authenticate an ssh connection. This is a security control that mitigates against private key compromise. This is great when you only need to execute a single command. But if you need to perform multiple tasks, you would normally need to create and install a separate key pair for each command, or just not bother making use of forced commands and allow the key to be used to execute any command.

from="<ip>",command="/usr/local/sbin/validate-rsync" ssh-dss AAAAZ5Hbl......

echo "from="${MYIP%% *}",no-port-forwarding,no-X11-forwarding,no-agent-forwarding,no-pty,command="rsync ${SSH_ORIGINAL_COMMAND#* }" $(ssh-keygen -yf ~/.ssh/rsync_rsa)" | ssh targetserver "cat - >>~/.ssh/authorized_keys" Note that the ‘command=’ restriction (http://larstobi.blogspot.ch/2011/01/restrict-ssh-access-to-one-command-but.html) will not apply if ‘/etc/sshd_config’ has already a ‘ForceCommand’ directive.