Francesco Pistillo @wildspecial

Chief Architect @ SAP ex Sr. Solutions Architect @ adidas Strategic Technologist, CTO, Product Obsessed and Business & Technology blender - Opinions my own N 40°26' 0'' / W 79°46' 0'' Joined October 2007

Tweets

3K
Followers

921
Following

4K
Likes

166

Francesco Pistillo @wildspecial

2 days ago

Can you please bake your statement with some data please? What is the methodology you used to keep the “entropy” and drift of the original goal intact and how did you measure that? Having your subagents working for hours without any human supervision and just “praying” LLMs will do the final magic sees a bit too naive IMHO… unless you used a methodology that I’m keen to learn 😅 Ps: the delegation to a subagent via n initial prompt, is, by a definition, a context loss that hopefully the subagents will try to “rebuild” if something is missing at “delegation time” Just one of a few papers on the literature: arxiv.org/abs/2606.08162

0 0 1 84 0

View Details

Francesco Pistillo @wildspecial

a week ago

@matteocollina @NVIDIAAI Why don’t you buy a small private data center? Jokes aside, most of your use cases can be achieved by AI inference providers with ZDR… local AI will be for sure a thing for 2027 tho 🤪

0 0 0 280 0

View Details

Francesco Pistillo @wildspecial

a week ago

100% ! But please… I think you forgot to put which hardware the whole thing needs to run. Ideally, other than the new Mac mini, we should have somewhere a consolidate list of “linux boxes” / power stations… ranked by power score (ram/compute etc.) , ergonomic score (mobile vs desktop vs mini) and a bunch of other kpis. Or ask @grok as usual 🤪😅

0 0 0 311 0

View Details

Francesco Pistillo @wildspecial

a week ago

Nope, you missed the point. For self hosting I meant all the cloud software we used “rent”. Productivity software (office suite, mail, project mgmt., notion-like software) , private chats (buzz or mattermost) … basically you need a power station where you self host all the software you need + a small quantized model (today) in future frontier-equivalent LLMs locally in conjunction with all the software you need for maximum privacy

0 0 0 19 0

View Details

Francesco Pistillo @wildspecial

3 weeks ago

@49agents Totally agree 👍

0 0 0 11 0

View Details

Francesco Pistillo @wildspecial

3 weeks ago

Don't trust AI Evals, build your own personal eval! Been using Opus/SOL level models for coding tasks: slow, not very accurate sometimes etc. One may argue and say: Skill issue, harness issue, context issue, memory issue... but then you try GLM 5.2 model and you become: faster and more efficient (reusing the same harness/context/skills....) Build your own eval!

2 0 2 1K 0

View Details

Francesco Pistillo @wildspecial

3 weeks ago

@Zai_org Guys, other than Evals that i don't trust anymore, you can simply say: try our model and you'll be shocked! :p Great work!! x.com/wildspecial/st…

Francesco Pistillo @wildspecial

3 weeks ago

2 0 2 1K 0

0 0 1 963 0

View Details

Francesco Pistillo @wildspecial

3 weeks ago

and i just discovered that z.ai released GLM-5.3 🤯 x.com/Zai_org/status…

Z.ai @Zai_org

3 weeks ago

Introducing GLM-5.3: Built to Code. Ready for Cyber Defense. - Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model - A major leap in cybersecurity, setting a new standard among open models Tech Blog: z.ai/blog/glm-5.3

951 2K 19K 5.9M 3K

0 0 1 75 0

View Details

Francesco Pistillo @wildspecial

3 weeks ago

Your are missing other factors: speed and throughput. If those beast generate and very short time thousands of line of code, as humans, we don’t have the bandwidth to digest that amount if we want to keep the promos that now with AI we are more productive and faster…. So sometime it’s convenience!

0 0 1 173 0

View Details

Francesco Pistillo @wildspecial

3 weeks ago

@NetworkChuck For power user like me, my friend…. The thing is that, normies don’t care what you care…. They don’t want to switch mode, they don’t care about control, they don’t care about privacy…. Think of your mom (if she is not in the tech industry) … will she use Hermes or grokbot?

1 0 0 413 0

View Details

Francesco Pistillo @wildspecial

4 weeks ago

Agreed! Who says that are mostly developers/ tech dude in the valley (I guess/suspect) Of course Lovable in the vibe coding space is not special or can be easily replicated, the thing at this point is “timing” and “traction”. If you are one of the first that puts the pierces puzzles together and you create a brand a first traction the moat is the “perfect timing” not much what it actually Lovable provides… and Lovable has gained momentum/popularity…. The only risk are some “moves” from consolidated AI labs (without naming them 😅) x.com/wildspecial/st…

Francesco Pistillo @wildspecial

4 weeks ago

@kylegawley You are forgetting the hybrid approach and “man/businesses in the middle”. So the statement is still true but you are using the wrong stack, that for developers, for business users companies such as #Lovable or #Replit will do the trick… so people will still build their own

0 0 1 452 0

0 0 0 227 0

View Details

Francesco Pistillo @wildspecial

4 weeks ago

You are forgetting the hybrid approach and “man/businesses in the middle”. So the statement is still true but you are using the wrong stack, that for developers, for business users companies such as #Lovable or #Replit will do the trick… so people will still build their own apps.. but with the right AI stack…

0 0 1 452 0

View Details

Francesco Pistillo @wildspecial

4 weeks ago

Computer use is the wrong paradigm for agents IHMO… it’s just a workaround for AI agents to “use” a computer “designed” for Humans… Instead of clicking here and there, agents needs to programmatically access everything.. WebMCP is the right direction for interacting with a website…. We should have something similar for desktop even tho the old AppleScript is something close to this paradigm…

0 0 0 24 0

View Details

Francesco Pistillo @wildspecial

4 weeks ago

Hey @levelsio we (some of us) know how magic LLMs are when you throw away unstructured data and start asking some clarity, I also did those exercises but please, try to advocate those having ZDR (zero data retention) policy on the inference providers (there are a few) and using some capable open models such as GLM 2.5 or Kimi because normal users get min 30 days data rententipn policy with frontier labs (codex/claude) and if some dude with a capable model will hack those labs (or their own lab models will escape the sandbox) your “very sensitive” data isn’t going to end well….

@levelsio @levelsio

4 weeks ago

A very fun exercise is to export all your bank accounts (as CSV) both personal and business and stock broker And dump it all into Claude Code and ask it to make a little dashboard showing your spending It's hard to know how much you actually spend over lots of accounts so this

114 41 1K 176K 1K

0 0 4 116 0

View Details

Francesco Pistillo @wildspecial

4 weeks ago

This is exactly what happened in my mind in 2022 when I realized the trend and potential of AI… the only way to keep a bit of human-driven touch for the internet is to find a way to let only human-verified content be certified … if a company/startup finds a way to do that… well it could become the new oil … certified human content!

0 0 0 70 0

View Details

Francesco Pistillo @wildspecial

a month ago

I’m curios to know, if you gave 1M context in one go, next turn hit immediately a summarization / compaction hence from step one you already loss information 🤪. Jokes aside, I’m wondering what do you think about context… it seems really the bottleneck people are trying to overcome with solutions like memory, context mgmt etc… do you think in a couple of months /year we are going to have 10M, 100M, 1K (like it happened with the RAM?) or we will go towards other directions?

0 0 0 1K 0

View Details

Francesco Pistillo @wildspecial

a month ago

The beauty now is that I look more social when I’m in metro and I talk to my agents. No voice to record and send back and forth! A real conversation in realtime with my friends (my agents) with a simple web app exposed via Tailscale (yes conversation continues also with lock screen) and with smart/semantic stop detection, interruptions, background agents pulling data from here and there while I talk with my main buddy … in what world we are living in…. Yes I have also human friends (still…😅)