sheepscreek

I'm speculating out on a limb here, but maybe the Mythos zero-days that were developed into working exploits involved a few of those "usually won't cause real-world problems" happenstances.

A model with all the world's knowledge might know exactly how to weave them together into a successful exploit. And if tomorrow's models are even more capable, perhaps what we're experiencing now is just a fire drill for what's coming.

Thanks for sharing this - I would not have looked into it otherwise. Apparently model retention is ~98% of the original at IQ1_M if I read that right. Crazy. Excited to try this now.

This is the point where you could ask an agent to unblock you. Don’t even need a subscription. Download opencode and pick the GLM-5.3-Flash (Free) model. Just be mindful to not use it on anything you don’t want to end up in training.

And like the author, are you suggesting that is because JB left the company?

Why would him being there have changed anything? What ‘hit product’ could JB have launched that you would have bought given everything that happened with Tesla.

This is an opinion piece - there’s no factual basis (nor any facts quoted) to support this narrative at all. Plus, key things missing from the calculus include:

- COVID, that put a strain on manufacturing during lockdowns and the eventual worldwide supply chain disruptions

- the new administration and its anti EV stance

- emergence of AI

- betting the future of Tesla on Optimus (product) and Robotaxi (service)

How much could JB Staubel have influenced Tesla during these times is anybody’s guess.

Yeah, I went down the rabbit hole scratching my head the whole time. This is about the group at Microsoft Research that was disbanded or reorged into what is now Microsoft Research.

To the author: 100% need this kind of thinking. It’s a balancing act. Historically though, dedicated labs haven’t had the best track record at commercializing their tech. The skillset differs vastly from running a product-focused business.

That said, even within this narrow scope - there are areas ripe for research.

Some of my particular interests include:

(1) Latent thinking/reasoning, which could arguably be more efficient but also opaque is one

(2) Separation of “intelligence” from knowledge - ie. keep the ability to reason like a frontier model and make tool calls (calling the ‘oracle’ or searching the web) but remove all parameters associated with niche knowledge/facts.

These are a couple off the top of my head that are very much aligned to scaling.

> If you actually recreated entirely new assets then it would get interesting. I suppose we will get there, and perhaps sooner rather than later.

Second that, and we'd get higher quality assets. Game engines could be rewritten/rewired to support ray-tracing and what not. Exciting.

Playing in freeplay, I feel like I am Jason Dessen from Dark Matter. So many green Nissan 350Zs. It is impossible to drive a 100ft without hitting someone like me.

Not just you. With a fourth of the world in an actual war, it doesn’t sit well with many of us.

That said, whoever made it might not be thinking about that. A kid could vibecode a game these days. I wouldn’t hold it against them.

Moreover, many AAA games glamorize war (Call of Duty, Far Cry, Battlefield…) or crime (GTA, Red Dead Redemption, Payday, Hitman…).

Ah that would explain a lot. Thanks.

To err is human. There’s no escaping it, I’m afraid.

^ This is basically what I was getting at. Thanks for putting it so clearly mate.

> Try to keep context shoved in a conversation thread under a token limit to save tokens.

That is another solid tip. If using subagents, use the main thread for only the conversation + launching subagents. Tell it explicitly not to implement anything, and only support integration efforts.

Doing so keeps the worker subagent's context clean - so it can stay laser focused on the task.

First off, I cannot pretend I have this solved. But I can share one thing that has worked for me, along with codex maintaining it's own local memory (turn it on).

Have your agent maintain a task journal. It is like a more detailed commit history, that includes tribal knowledge gained during the task and additional context, eg.

    X is incorrect but we cannot change it due to some other service depending
    on it, don't use Y as it is deprecated in favor of Z.

Reset your context frequently instead of only relying on compaction. Use the following prompt to summarize the current context for the new agent:

    Describe the current project or larger task in progress, then give a 
    breakdown of all the planned sub-tasks with status, explain any learnings,
    then note what work is remaining. Write this in a file called HANDOFF.md

Give it a try and feel free to share any improvements.

I should have clarified I meant XML is no longer seriously considered for any new protocols/formats.

For XML, we can add docx/xlsx/pptx/odt/ods/odp to the list.

I have good memories of using XSLT. It was a neat idea at the time.

I suppose the world has nearly moved away from XML (for browsers, not yet for folks still enjoying SOAP APIs). And as a result, XSLT has fallen into obscurity.

You should seriously take your business somewhere else. Vibe-coding is the norm now and Replit does not add any meaningful value anymore.

Very cool stuff. Thanks for sharing.

Looking at the images, it strangely appears as a modern second tier (ie. not NY or London), post-war western society. More industrial and progressive than I would have imagined.

Yep, this is one of the the darkest patterns that I've seen in a Google product. No paid tier can help you get rid of it.

The only way out is to stop being a 'consumer', and opt for a business account.

> The “just by looking at the raw weights” setting (raw floats in context, no tools, no running the network) only appears in experiments 1, 2 and 4, on networks of 17K–172K weights.

The author is being equally disingenuous.

That piece by itself is insightful. Meaning, LLMs are capable of deciphering what a model can do by looking at the weights alone.

The fact that we can’t squeeze enough of them through the context is a technical limitation. However, this does not diminish the importance of this finding/discovery. This is an unexpected emergent behaviour.

V100 is a beast if you feed it the right format. A lot of these older cards don’t do well with lower quants. So instead of using Q4, try Q8 or FP16.

If you’re still getting poor performance, point an Astra agent at it and ask it to benchmark and find the pressure points.

Also speculative decoding has changed the game completely. I’ve managed to 2-3x’d the theoretical max for my setup. Apple Silicon is a bit more limited here, and cannot deal with large DFlash block sizes. But AMD and Intel chips are game.

Good luck and godspeed!

> I have not seen an exploit in the wild...but the risk is blown way out of proportion

That's a very naive view on it (now) and I'll try my best to explain why.

The bad actors can discover multiple 0-days on the fly to gain access to an employer/company infrastructure with valuable data/information, or in the case of Binance, crypto. They could chain multiple such attacks going from one-company/provider to another, to another, to eventually reach their target.

To anyone interested, this is good reading material on the topic:

(1) study the Binance hack in detail

(2) OpenAI huggingface hack - focus on 0-day chaining

Honestly that shouldn’t even matter anymore.

The things the original commenter wants to know are exactly the things that will count in an interview. The reasons behind the decisions. Not the actual implementation.

Implementing anything, even something of significant complexity is trivial with frontier agents now. So build your experience / knowledge in a technology-agnostic way. The language/app/stack is superficial.

Edit: Speaking anecdotally; it’s what I’d look for in a candidate if I was hiring today

The original tui-app is more stable. I’ve had instances of my agents randomly getting killed on herdr-core + macOS app stack.

Wow that is incredible. I believe the colors are overlaid (or underlaid) to help with illustration but the position/placement is a real photograph.

Integrated experience. You can connect to multiple machines and access terminals on them through spaces in the left nav. Agents across all spaces are listed in a dedicated section for quick access.

Each space is persistent contains persistent terminals like tmux, so disconnections do not affect them.

We don’t know if the person did any such thing (bad mouth the company or its products).

However, in the scenario they did, then the company would have grounds to fire them.

She approached to ask for permission. If they had any reservations, that was the time to discuss them. She did what was expected of an employee. The fact that they blessed her and then went back on their word seems wrong.

Not a lawyer but my intuition says there might be some legal grounds here favouring the wife.

My take is someone even higher up found out later and did not like it. Now it’s entirely possible the individuals who gave their blessings were supportive of the party the wife was standing for, and the ones above them weren’t? Still whoever blessed the campaign did so in their capacity as a senior officer of the company. A total shit show nevertheless.

To the original commenter: sorry this happened to your wife. Hope she’s able to find another job that treats her right. Also congrats on the baby, hoping parenthood makes up for all of this and more. Sending the best of wishes your way.

I went by their illustration of GPT 6 Astra that looks like a spiral galaxy[1].

[1]: https://openai.com/index/gpt-6-astra/

Hmm..they did come up with Bard.

Anytime I see one I can’t shake off the feeling I got when I first saw Watchmen. I thought “this is so dystopian and bizarre; so out there”. Didn’t have this in my bingo card for the decade.