Jeremy Bayley

I tell stories and build relationships with people and businesses.

We are living in a time of awsome victory and it is exciting! Eat cake and celebrate!

[Jeremy Bayley @Jeremy_Bayley]

Bro, Grok is hardly used outside of Twitter comments. Thats why 4/5ths of Collosus 1 is being rented out to Anthropic while Collosus 2 is used for training. Grok is not a great model. And using AI is not fraud, if it solved the bounty the payment was for, that means the bounty was solved.

[Jeremy Bayley @Jeremy_Bayley]

@ihirwart @sama Its not a charity. What your using is a capped for-profit subsidiary. The non-profit oversees the structure, it does not provide the service. This split occurred in 2019.

[Jeremy Bayley @Jeremy_Bayley]

Having used the Claude 20X plan, Claude entry level plan, and the OpenAI entry level and the $200 plan, the two are quite comparable. For my software development work i generaly find Codex to be better beacuse the inference is generaly faster than Opus, the agent harness has fewer bugs, and it generates code that compiles more consistently than Opus.

If you followed the Claude Code harness source leak, the harness is one of the worst performing on the market. It has long been suspected that anthropic used a slightly different model for API service vs the subscription due to the API performing better, turns out they were shooting themselves in the foot with the harness.

[Jeremy Bayley @Jeremy_Bayley]

@goodhunt Why is it using the Nvidia name and logo? This is not a Nvidia repo. The real one is https://t.co/N8ua5RIwsf

[Jeremy Bayley @Jeremy_Bayley]

@DAIEvolutionHub @grok the PDF is freely availible at https://t.co/OYgB73ZRg3 , why would the OP want to DM users to send them a private link instead of providing the freely availible public link to the original document instead?

[Jeremy Bayley @Jeremy_Bayley]

@RippleXrpie Neither Canada, USA, or the EU import any oil from Iran, so Iran oil prices are mostly irrelevant to the target audience of their piblic statment.

[Jeremy Bayley @Jeremy_Bayley]

@gator_gum @__Buu__ @elonmusk

[Jeremy Bayley @Jeremy_Bayley]

@gator_gum @__Buu__ @elonmusk I can confirm that it is true, go to https://t.co/oopv8Km796 then go to the search bar where it says "Search Full Epstein Library"

[Jeremy Bayley @Jeremy_Bayley]

@levelsio It is impossible to aquire it, its not a commercial product and its voluntary developed. They could fork it, provided they retain the original license notice, and continue developing it on their own. Anthropic being Apple-like & anti-competition wouldn't like the association.

[Jeremy Bayley @Jeremy_Bayley]

@Jason @OpenAI @xai @Gemini @claudeai Your company was only spending $10k yr? Why didnt you just keep it and expand utilization accross other models? Don't let those tokens be wasted, a single person running codex can use all of that $10k to great effect.

[Jeremy Bayley @Jeremy_Bayley]

@sportvilde @JulianGoldieSEO Your 12B model is likely around 8-10GB, the smallest Q1 quant of minimax 2.1 is 55GB. So you need at least 5x more VRAM.

[Jeremy Bayley @Jeremy_Bayley]

@ItIsWha92763579 @JulianGoldieSEO 96GB for Q1. 64GB VRAM if you offload some to RAM. Bit you dont want to run that small of a quant. Q6 with 200GB VRAM is more where you should be at a minimum. Keep in mind the comparisons for benchmark are likely BF16.

[Jeremy Bayley @Jeremy_Bayley]

"ever", You sure about that? We havent reached the plateau yet and ever is a long time. Look where things were 1 year ago compared to now. Sonnet 4.5 right now isnt even the leader and its close to being matched by an open 200GB (Unsloth Q6 UD, minimum it should be run at) model. Sonnet 4.5 will be the bottom of the barel by end of 2026 and probably not even availible via Anthropics subscription anymore.

[Jeremy Bayley @Jeremy_Bayley]

@davidofug @JulianGoldieSEO It does have an integration per-say. Its just a model. You need to hold it through somthing like vLLM or Olama, then point a coding agent like roo or kilo (vscode plugins) to the local address.

[Jeremy Bayley @Jeremy_Bayley]

@davidofug @JulianGoldieSEO No,this model requires at least 200GB of VRAM / Allocated Unified to be useful.

[Jeremy Bayley @Jeremy_Bayley]

@religAIon @JulianGoldieSEO Show me a phone that matches the performance of a desktop with a cross compatible desktop mode and then we can talk

[Jeremy Bayley @Jeremy_Bayley]

@Purring_Lynx @liquidai Maby read the model details next time that explains what they are talking about? Did you train it? Did you use it for the kind of knowlage task they are referring to? Did you note the part where it says "We do not recommend using them for tasks that are knowledge-intensive..."?

[Jeremy Bayley @Jeremy_Bayley]

@Purring_Lynx @liquidai Why are you posting results of an undefined test for knowlage on a small model? Knowlage is not what these models are for. Use the correct tool for the job.

[Jeremy Bayley @Jeremy_Bayley]

@MaxtheCrowhouse @theamelia__ It helps to know what the definition of the word is, and the population statistics of where its being applied.

[Jeremy Bayley @Jeremy_Bayley]

@MrDrainPipe @corbin_braun It depends on the license. All the major coding agents wave the rights to the user. Many of the open models provide permissive license to output. Some licenses have conditions such as a royalty if the output is used in a product that grosses x dollars yearly.