Billovskii

"AI explorer | Local LLM wrangler | Coding wannabe | Chasing xAI dreams"

Grok-It is in | 𝙶̶𝚘̶𝚘̶𝚐̶𝚕̶𝚎̶ ̶𝚒̶𝚝̶ is out

[Billovskii @TheArcAge]

@ashxhart @b_nnett TensorFold makes all the difference between barely usable and extremely capable and a pleasure to use. All the difference in the world. Thanks Ash.

[Billovskii @TheArcAge]

@NixFred What a great shirt Fred and what a great son your dad has! 👋

[Billovskii @TheArcAge]

@DaveShapi This is a must read. Do not fail to read the last 3 paragraphs from Dave. In the human strength category (next to last paragraph) I would add taste, culture, and knowing "what is worth doing". Great essay!

[Billovskii @TheArcAge]

@IndependentEco @NanneWielinga I'm with Eco here. Claude has the ability, in a long chat back and forth, to almost become like a trusted colleague where the end result is larger and better than than Claude or me alone. Doesn't happen all the time, but there are those "ahha" moments occasionally with Claude.

[Billovskii @TheArcAge]

@TheGuruOfYou @IndependentEco Omars... A great synopsis of the strengths/weaknesses of ChatGPT through Gemini in short, easy to understand, terms for each. Great job!

[Billovskii @TheArcAge]

The chat box era worked with low dollar subscriptions because the process is one-on-one interactive; it takes time to type the prompt and time to read the response back and forth a few times. Starting with just one agent, time for agents to read and write responses are almost zero but each loop uses tokens just the same and they do it over and over until the task is completed. Then they move on to the next task. Low dollar subscriptions get blown out in hours, not days.

Now add swarms of agents starting with about ten. Then add fast mode to get results quicker and the results get out of budget quickly. I wondered what all the 1+ Trillion $$ gigawatts of computer power that is being built in 2027 was supposed to do. I actually now believe that it is being built to deliver enough compute to accommodate agent swarms. Now who can afford all of that is another matter.

[Billovskii @TheArcAge]

@buildwithhassan This is scary and adds so much uncertainty in evaluating models. It is almost saying to me find a model, find the right harness, and don't change a thing. In any case, it add complexity to evaluation of models and their potential.

[Billovskii @TheArcAge]

@MichaelHyatt @Apple @grok @Muse Long time Apple fay-boy here, since Steve Jobs that's how long, and the new Siri is slow and unreliable. I had hoped for lots more, but mostly lost faith in anything Apple.

[Billovskii @TheArcAge]

@NotJustShaking This will great to find out. If it works I have 2 or 3 like it sitting around in a corner somewhere around here.

[Billovskii @TheArcAge]

@martiano M4 Max, long time Apple fan-boy, is waiting eagerly and patiently (well maybe not so patiently) for this to happen. Love Omarchy on my old mini pc, but holding out the real love for when it hits my M4. Kudos to the team for all the work on this.

[Billovskii @TheArcAge]

@kimmonismus This is quite disturbing to me but it appearing that local source, open weight, open source models, which principally have come from China, are losing ground to the frontier models and that's now a good sign. It makes investing in local hardware harder to defend for sure.

[Billovskii @TheArcAge]

At current API prices their margins to deliver "inference" for chat, cowork and code have to be above 80% so they are definitely heading in that direction. I expect the following levels of subscriptions; $100, $200, $500 & $1000. I expect them to deliver $500 in equivalent API pricing for the $100 sub and maybe up to 20x that for the $1000 subscription.

I think the free ride on usage for the low tier subscriptions will be greatly depreciated. I hope I'm wrong, but with a projected 1.4 - 2.0 Trillion dollar valuation, they have to deliver earnings.

[Billovskii @TheArcAge]

I'm with Natanael. Argon's writing is top notch almost best in class but it too seems lazy. What do I mean by lazy? Lazy in not quite finishing the full thought and/or sum up at the end. Claude and chatGPT both seem to "finish the chat" while Argon in all of my use thus finishes with me always wanting a little more of the chat discussion. Plus the follow up suggestions are always really specific and don't lead to follow up on their own. I usually have to come up with my followup ideas independently.

The tone, the writing "voice", the "aha moments" that come from Argon are some of the best in frontier models that I have seen in quite some time. I don't leave a conversation trying to know what Argon is trying to say. Jargon and acronyms are at a minimum which is really good.

To try and put the "hurry" another way, it seems Argon is always looking to move on to something else.

[Billovskii @TheArcAge]

@edo_lunardi Not going to wait for later, THANK YOU!

[Billovskii @TheArcAge]

@alexhillman Alex... thanks for posting this. Looks like the one for me and I never heard about it until your post here. Kudos for the find.

[Billovskii @TheArcAge]

@machinegenie @seesa_ai What are the levels of equipment starting with the RTX6000 pro's and up a notch from there. That's a potential for my company that I am at least considering.

[Billovskii @TheArcAge]

@karpathy 1. The impact scales with ambition, problem size, and horizon.
2. You need a reservoir of big, difficult problems that you really care about.

These are two important points in Karpathy's article and neither touch the 99% of the normal chatbot, google replacement public users.

[Billovskii @TheArcAge]

Prior to TensorFold and @MiaAI_lab recipes it was getting enough TTFT and speed (tok/sec) to levels that made a chat interface useful. That is so much better that it no longer is any hindrance of using local model. Now it is getting models large enough and accurate enough to replace not Fable level intelligence but Sonnet 5.5. Local needs slightly more intelligent models that can utilize yours and Mia's "magic" for 95% of our agent's tasks.

It seems the trend is towards making local model work is smaller and smaller gb configurations. For me personally, I want larger, more post-trained, company specific models or at least the ability to turn GLM-5.3 level models to more personalized need models but sovereign oriented.

[Billovskii @TheArcAge]

@ryanrhughes @jondkinney @tiredkebab @maralcbr Does Omarchy on the M3, M4 and M5 chips run dual boot or will it run Omarchy as the only operating system on those hardwares?