Mike Leiter @MikeLeiterDev

Amazing results! DeepSeek latest v4.1 1017 (PLEASE stop versioning with dates!) at over 110 tokens/second? Of course, it better be amazing results, for $10k in hardware! 😊

BlackwellBoy @Blackwellboy

13 hours ago

BLACKFOLD IS LIVE. 🤯 Uncensored DeepSeek-V4.1-Flash 1017 running locally on two DGX Sparks. No restrictions. Blazing speeds: ⚡️ 2,008 tok/s prefill (128K context) 💻 113 tok/s code decode Virtually zero refusals matching api quality. Everything I fixed to get there 🧵

19 12 105 60K 72

Typically, I would start with just a prompt. The prompt directs the Agent to do some part of the workflow. If it does it okay, then I'll turn it into a skill. Just a markdown skill. I'll probably tell my agent to write the skill, based on the work and prompts we just had in the last session. Then comes the tweaking. Every time you get to that part of the workflow, use the skill instead of prompting. Update the skill when it doesn't perform. Put in the skill for the agent to run checklists of the output, etc. Once that individual step is reliable, add the next one. Then start chaining them together...

Most people learn AI backwards - they chase tools and prompts. What if the problem isn't the tool or what you learn, but the order? Here's how I'd learn AI from scratch, if I had to start today. x.com/i/article/2112…

@Blackwellboy Awesome report. You keep crushing it!

Amazing investigation. I did *NOT* expect to see Grok Build tied for 1st in a test like this - good for xAI! Pi agent isn't a surprise though. Many times less is more, you know?

Paweł Huryn @PawelHuryn

23 hours ago

I repeated DeepSeek V4.1 Flash 1017 benchmark 89 times across 9 harnesses to understand which works best. High effort, 5 attempts each: Pi harness: 40.4 Grok Build CLI: 40.4 DeekSeek: 39.2 OpenCode 2.0: 38.6 OpenCode: 36.6 Claude Code: 36.4 Codex: 33.2 Hermes Agent: 32 Kilo

150 77 1K 89K 466

What is up with X's algorithm for your feed? It seems you only get people who you have fairly recently liked or engaged with. After a couple months my feed became static - it was always the same 20 or 30 people's latest post taking up all the spots. It is just me this is happening to?

@PawelHuryn And... at 150 times lower cost than the current leader, Sonnet 5.5.

More Shadow-ban Saga, day 3: Today I got hit with a "we've seen unusual account activity, verify your email" from the Great Enforcer In The Sky, with my account presumably locked until I did. Which I immediately did. Upon connecting back to my account, I still have all my followers, but my profile is showing that I'm following 0 people now. Hopefully that resets? I thought this kind of B.S. disappeared with Elon, but now with all the X automated bot protections - run mostly by X's own automated bots - things seem to have gone off the desired track a bit. The bots seem to have gone a little wild! There's a hot new program for an up-and-coming streamer/influencer - "AI Bots Gone Wild!!!"

@PawelHuryn I think they get an "A" for effort. Meanwhile, I find the last 3 lines for Luna-6 max, Haiku 5.5 xhigh and Luna-6 xhigh quite a fascinating set of entries! Clear tradeoffs, for sure.

@_ryu15_ Ouch! -24 (that's a minus in front) on AA Omniscience bench. Careful trusting anything it tells you... Just as bad as Qwen3.6-35B/A3B though. And it should be even faster!

@morganlinton Wow - that's not even the Full version. Same API cost as Opus 5.5 though, right?

Looks like the joke is on me. Somehow the cancellation (which you get no transaction # for or email to confirm) was undone and my card was charged today! For an account that's unusable for normal purposes. Thanks for nothing, X. I tried to remove my card after cancelling so this WOULD NOT happen, but they wouldn't let me since "I have an active subscription using that card". No I didn't. I cancelled it. Now I've cancelled it again. Still can't remove the card. What a completely POS design. They couldn't care less. Sorry about the venting. I suppose it's natural. Now I get to waste time challenging the card with the automated systems of my credit card company. Joy. Looks like all future purchases of every subscription need to have their own virtual card that I can shut off at a moment's notice. Ironic that it was an AI company that hardcoded the behavior that bit me, rather than an AI Agent.

Shadow-ban saga, Day 3: It appears views are allowed to trickle in. One of my posts had 7 impressions last I checked. I'm wondering how many of those impressions were mine. But I did get a like and a reply or two, so some of you can see some of these!! 😆 Thankfully this hit with only 3 days left in my monthly Premium+ subscription, so I cancelled - I'm not putting good money after bad. But that does mean I will lose some capabilities and presumably reach. Most notably the "long post" feature. That's going to really suck rocks, lol! 😯😆 Thanks to all who support me who have found my engagement to have some kind of value!! I love you all! ♥️

@Atifel_ Oh - that's a lovely side-effect! 😊 Thanks for the info!

Daily Shadow-Ban Status: Day 2 Please reply or DM me if you can see this! Thank you!!

Anthropic Subs are better than OpenAI - but only if you use Haiku!! Luna and Haiku are about the same quality vs. cost per task - when looking at API costs, if shifted by a rung of effort or two (Luna max is about the same as Haiku high). And there's a bit of a consideration that Haiku becomes 5 times more expensive if you go over 100k context size. But, at least for smaller, shorter jobs, Haiku looks very solid for roughly the same cost. But since Anthropic gives 40-45 times subsidy for their subscriptions, while OpenAI only gives 10-12 (after recent updates). So, while Opus 5.5 is 3-4 times more expensive than Sol 6.1, make their sub usage about the same for a similar task, the same is not true for Haiku vs. Luna. Ignoring the 100k context cliff, you will get a lot more total usage from Haiku than you do with Luna for the same type/size of sub. You can also run Haiku at lower settings for the same quality result, which will be faster - or run it at higher settings and still be less usage than Luna on the OpenAI sub. So, if you mix in Luna/Haiku for lower end tasks at all on a subscription, Anthropic seems to be the higher value right now. Of course, tomorrow a new model could release and change the calculus. We're probably all well aware of that problem by now!

5/10. It's difficult to manage conversations with the bots. 1. No easy way to carry on multiple conversations with a given bot vs. others. Just unnesting related conversations with a given bot, for example, under that bot in the bot window might be a simple fix. 2. For multi-node, it's just not there yet. I have 4 nodes (plus the phone, when the apps drop), and getting them to connect, not step on each other, etc. is difficult, even with just one main node and the rest just having one local bot to handle admin on that PC. Having to swap nodes to get a view for just that node is confusing and doesn't give the expected results (or there are still gaps here). 3. I don't understand how the layout is supposed to work - maybe it's my multi-node usage. I should be able to see files in a working directory and pull those up easily, for example, with Terminal access easier than it is as well.

Anthropic pulls even in sub value, comparing Opus 5.5 to GPT 6.1 Sol! According to several benchmark tests, Opus 5.5 is roughly 3-4 times the cost of GPT 6.1 Sol - if you are comparing API costs and looking at cost per task in various benchmarks. But, Anthropic gives about a 40-45 times multiplier for "API value" when subscribing - at least for the 20x/$200 plan. This has been calculated by multiple sources I've seen now, and accounts for OpenAI's recent change to reduce how much subsidy their accounts have compared to API. Meanwhile, OpenAI is only giving 10-12 times API cost for Subs. When you put those two together, getting an account at OpenAI and using Sol 6.1 all the time vs. getting an Anthropic account and using Opus 5.5 all the time will allow roughly the same tasks to be completed from a usage standpoint. This doesn't factor in those paying API costs or credits. There if Sol 6.1 will do, you can get a 3-4 times advantage of Opus 5.5 from a cost standpoint. Now - a Haiku vs. Luna is another situation...

Okay, 5 impressions in 4 hours is meager, at best. Maybe it just lasted a day or two. Need some real content to post now to test it out...

@MichaelGannotti @aionaedge Nice! Data from Experiments is always great!