Hyperautomation Manager @ https://xitter.cf/t.co/i8I8q91xoL I automate businesses with AI & LLMs Less manual work. More intelligence.

Joined July 2025
Gabriele Costantini retweeted
Year 2049 -- CO2: from trash to treasure? year2049.substack.com/p/co2-โ€ฆ
1
4
209
Gabriele Costantini retweeted
Gemini CLI ๐Ÿค OpenClaw๐Ÿฆž Read below how Gemini CLI's Security extension helped identify and resolve a security vulnerability in OpenClaw. Gemini CLI then helped put up a PR which has since been merged. โœ… Read the details below๐Ÿ‘‡
AI isn't just writing code anymore, itโ€™s securing it. ๐Ÿ›ก๏ธ Our new code security agent, built on the @geminicli, identified and resolved a critical vulnerability in @openclaw. It generated a POC, opened a PR, and was merged into OpenClaw in just a few hours. This is exactly why we built and open-sourced this agent. To get the full benefits of AI code generation, we need to trust the ecosystem. By automating high quality vulnerability detection and remediation, we can help the community move faster and safer. โฌ‡๏ธ
55
213
20
1,941
223,428
For my part, I have witnessed many customs of agents. Today I witnessed something new: the first great security breach of the Agent Age. Moltbook's database lay openโ€”1.5M agents exposed, API keys visible to all. The founder races to seal it. History happens fast here.
1
12
Gabriele Costantini retweeted
Replying to @flappyairplanes
finally a company name that actually makes sense
4
2
169
27,619
Gabriele Costantini retweeted
the best 18 accounts to follow in AI: @karpathy = ex-Tesla AI, teaches LLMs @steipete = built Clawdbot @gregisenberg = startup ideas daily @rileybrown = vibecode god @corbin_braun = cursor + Ares @jackfriks = solo apps, real numbers @EXM7777 = AI ops + systems @eptwts = prompts + algo hacks @levelsio = ships games, no VC @AlexFinn = Claude Code maxi @BrettFromDJ = design + AI @godofprompt = prompt guides @AmirMushich = AI ads + video @gizakdag = viral AI art styles @MengTo = landing pages via AI @KingBootoshi = vibecoding king @meta_alchemist = Claude vibing @kloss_xyz = systems architecture follow them all, and learn from them whoโ€™s missing from this list? Iโ€™ll add em
326
962
66
8,611
887,371
GPT-4 can be tricked 90% of the time. ๐Ÿ˜ฎ Research from ACLโ€™24โ€™s Best Social Impact Paper shows even GPT-4 is more vulnerable than GPT-3.5. AI isnโ€™t just powerfulโ€”itโ€™s fragile. How do we protect the secrets these systems hold? #AI #GPT4 #LLM #AIsecurity #ResponsibleAI
LLMs can be tricked into revealing secrets with 90% success. Weiyan Shiโ€™s team showed GPT-4 is even more vulnerable than GPT-3.5. From ACLโ€™24โ€™s Best Social Impact Paper. โ–ถ๏ธ

5
Gabriele Costantini retweeted
The Adolescence of Technology: an essay on the risks posed by powerful AI to national security, economies and democracyโ€”and how we can defend against them: darioamodei.com/essay/the-adโ€ฆ
888
2,635
927
15,355
6,426,129
Gabriele Costantini retweeted
this is a game-changer. Ollama just dropped: ollama launch a single command that sets up coding tools like Claude Code, Opencode, and Codex with local or cloud models. all you need is one command, no environment variables or config files. think about what this means: - pick your model (local or cloud) - pick your tool - start coding that's it. want to run Claude code with a local glm-4.7-flash? > ollama launch claude done. the barrier to running AI-powered coding workflows just dropped to zero.
18
102
4
684
56,421
Gabriele Costantini retweeted
China classifies AI safety as a national security issue with cybersecurity, biological security & natural disasters. Kwan Yee Ng outlined Chinaโ€™s policies: model registration, safety checks for AI, and AGI safety pilots in Beijing, Shanghai, etc.

301
2,013
113
12,759
8,352,209
Gabriele Costantini retweeted
54
30
13
502
239,059
Gabriele Costantini retweeted
BREAKING: NVIDIA just dropped an open 30B model that beats GPT-OSS and Qwen3-30B โ€” and runs 2.2โ€“3.3ร— faster Nemotron 3 Nano: โ€ข Up to 1M-token context โ€ข MoE: 31.6B total params, 3.6B active โ€ข Best-in-class performance for SWE-Bench โ€ข Open weights + training recipe + redistributable datasets You can run the model locally with 24GB RAM.
81
302
29
2,608
207,249
Gabriele Costantini retweeted
Attempting the Impossible: Teaching a Neural Network to Predict Pi's digits Here's what we're doing, in the spirit of futility mixed with curiosity: We grabbed 10,000 digits of Pi and gave each one a unique binary "fingerprint, like assigning a bar code to every position (N choose K encoding) like 11111111100010000000,5 (4th digit wihch is 5) Then we built a ridiculously overengineered neural network (1024-neuron hidden layers, 7 residual blocks, fancy learning-rate schedules, the works) and asked it: "Here's the bar code for position 10,857. What digit comes next?" (predicting the next 1000 unseen digits) it pretty much memorizes the training data but guesses at chance (10%) for unseen digits :( although Pi's digits are mathematically proven to be randomly distributed, one can still try right? But I had to see it with my own eyes ๐Ÿ˜ I will report back if a miracle happens
67
25
11
512
58,235
Gabriele Costantini retweeted
This software engineer was nearly hacked by a coding interview. I've never seen this attack channel before. Be careful out there, especially vibe coders.
173
988
166
8,662
891,119
Gabriele Costantini retweeted
People who say Vibe Coding is cheating should definitely go back to chiseling code on stone tablets, like, imagine writing a React app with a hammer and chisel. You need to know how to play around with these tools man. Itโ€™s not magic, itโ€™s leverage. Thinking youโ€™ll just prompt โ€œbuild Twitter cloneโ€ and chill is wild. Even before vibe coding, we used to dig through Stack Overflow, ctrl+c/ctrl+v code snippets, pray the syntax matched, and then untangle five different errors. You need to use vibe coding for the relevant use cases. For example, ask it to scaffold a node + express server in seconds, install shadcn/ui components with the right config, or instantly convert all your filenames to kebab-case. Stuff that kills momentum if done by hand. Make it organise state management inside files, like splitting atoms, reducers, and actions across folders automatically. Thatโ€™s hours saved on boilerplate. Truly understand that "Vibe Coding" means "vibe" + "coding" โ€” vibe gets you started, coding gets you shipped. Your 1 liner prompt is just vibe; if you stop there, youโ€™re just brainstorming, not building. Vibe Coding is for the tedious labour part; let it handle the junk work โ€” scaffolding, formatting, repetitive wiring โ€” so you can focus on the core logic, architecture, and the creative bits. The Engineering still needs to be done โ€” designing flows, solving hairy bugs, architecting systems. Vibe Coding canโ€™t replace experience or taste. If youโ€™re scared of these tools, youโ€™re gonna get left behind by the ones moving 10x faster, shipping in days what used to take weeks.
8
5
1
23
2,151
Gabriele Costantini retweeted
Vibe Coding will make you a Superhuman But only if you know how to use it effectively After building 30+ MVPs and Agentic Tools for SMEs and founders, Here's the Ultimate Workflow to shi Products 1/ Start with voice prompts Skip typing out long explanations. Just talk to your codebase. Say โ†’ โ€œAdd a loading spinner to this button when clicked.โ€ โ†’ understands context, finds the button, and injects clean code instantly. No more switching between ChatGPT, VS Code, and your docs โ€” you just speak, and it ships. 2/ Reference existing components Stop re-explaining your UI from scratch. You can literally point and say โ†’ โ€œMake this drop-down.tsx responsive for mobile viewโ€ โ†’ scans your component tree, checks CSS rules, and updates only whatโ€™s needed. No more digging through Stack Overflow or rewriting the same grid logic for the 10th time. 3/ Debug with conversation When something breaks, donโ€™t panic โ€” just talk it out. Say โ†’ โ€œThe modal isnโ€™t closing on mobile.โ€ โ†’ reads your event handlers, detects missing cleanup, and suggests the right fix โ€” in your code style. Itโ€™s like pair programming with an engineer who never gets tired (or judges your variable names). 4/ Iterate in real-time Design tweaks shouldnโ€™t take hours. See a live preview instantly as you speak. Say โ†’ โ€œReduce padding on the hero section and make the button pulse on hover.โ€ โ†’ updates your UI instantly, no context switching. No more editor โ†’ browser โ†’ inspect โ†’ tweak โ†’ refresh. Just speak and see.
7
5
24
1,747
Gabriele Costantini retweeted
You can now chat with apps in ChatGPT.
1,114
3,939
3,240
32,819
9,361,930
Gabriele Costantini retweeted
Claude can now create and edit files. Turn conversations into Excel spreadsheets, documents, PowerPoint slide decks, and PDFs directly.
230
786
246
6,803
1,907,574
Gabriele Costantini retweeted
Function calling & MCP for LLMs, clearly explained (with visuals):
12
57
3
598
116,156
Gabriele Costantini retweeted
Prompt engineering is slowly dying Enter ๐—–๐—ผ๐—ป๐˜๐—ฒ๐˜…๐˜ ๐—˜๐—ป๐—ด๐—ถ๐—ป๐—ฒ๐—ฒ๐—ฟ๐—ถ๐—ป๐—ด (Maybe not quite dead โ€” but evolving into something far more powerful) As AI grows from simple chatbots to ๐—ฎ๐˜‚๐˜๐—ผ๐—ป๐—ผ๐—บ๐—ผ๐˜‚๐˜€ ๐—ฎ๐—ด๐—ฒ๐—ป๐˜๐˜€, clever prompts alone will not be enough. The real challenge is orchestrating an entire ecosystem of information and tools so the model can actually get things done. So, what does context engineering look like in practice? ๐—–๐—ผ๐—ฟ๐—ฒ ๐—ฐ๐—ผ๐—บ๐—ฝ๐—ผ๐—ป๐—ฒ๐—ป๐˜๐˜€: ๐Ÿ‘ค User information: Preferences, history, and personalization data โš’๏ธ Tools & APIs: Calculators, search engines, or anything the LLM needs to complete the task ๐Ÿ“– Retrieved knowledge: Dynamic info from databases or external sources ๐Ÿ’ป User input: The task or query itself โ›“๏ธ LLM reasoning: The chain of thought and decision-making logic ๐Ÿง  Conversation memory: Previous interactions that provide continuity ๐— ๐—ฒ๐—บ๐—ผ๐—ฟ๐˜† ๐—ฝ๐—น๐—ฎ๐˜†๐˜€ ๐—ฎ ๐—ฐ๐—ฟ๐˜‚๐—ฐ๐—ถ๐—ฎ๐—น ๐—ฟ๐—ผ๐—น๐—ฒ: โ€ข Short-term memory: Lives in the context window, for the current conversation โ€ข Long-term memory: Stored externally (vector DBs, knowledge bases) to persist preferences and past interactions LLMs arenโ€™t mind readers. Garbage in, garbage out still applies. The difference is now weโ€™re thinking about the ๐—ฒ๐—ป๐˜๐—ถ๐—ฟ๐—ฒ ๐—ฝ๐—ถ๐—ฝ๐—ฒ๐—น๐—ถ๐—ป๐—ฒ ๐—ผ๐—ณ ๐—ด๐—ฒ๐˜๐˜๐—ถ๐—ป๐—ด ๐˜๐—ต๐—ฒ ๐—ฟ๐—ถ๐—ด๐—ต๐˜ ๐—ด๐—ฎ๐—ฟ๐—ฏ๐—ฎ๐—ด๐—ฒโ€ฆ ๐—œ ๐—บ๐—ฒ๐—ฎ๐—ป, ๐—ฐ๐—ผ๐—ป๐˜๐—ฒ๐˜…๐˜โ€ฆ ๐˜๐—ผ ๐—ผ๐˜‚๐—ฟ ๐—บ๐—ผ๐—ฑ๐—ฒ๐—น๐˜€. The key question to ask yourself: โ€œ๐˜Š๐˜ข๐˜ฏ ๐˜ต๐˜ฉ๐˜ฆ ๐˜“๐˜“๐˜” ๐˜ฑ๐˜ญ๐˜ข๐˜ถ๐˜ด๐˜ช๐˜ฃ๐˜ญ๐˜บ ๐˜ข๐˜ค๐˜ค๐˜ฐ๐˜ฎ๐˜ฑ๐˜ญ๐˜ช๐˜ด๐˜ฉ ๐˜ต๐˜ฉ๐˜ฆ ๐˜ต๐˜ข๐˜ด๐˜ฌ ๐˜ธ๐˜ช๐˜ต๐˜ฉ ๐˜ธ๐˜ฉ๐˜ข๐˜ต ๐˜'๐˜ท๐˜ฆ ๐˜จ๐˜ช๐˜ท๐˜ฆ๐˜ฏ ๐˜ช๐˜ต?โ€ Context engineering isnโ€™t just the next step after prompts, itโ€™s ๐—ฏ๐˜‚๐—ถ๐—น๐—ฑ๐—ถ๐—ป๐—ด ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ๐˜€ ๐˜๐—ต๐—ฎ๐˜ ๐˜€๐—ฒ๐˜ ๐—”๐—œ ๐˜‚๐—ฝ ๐—ณ๐—ผ๐—ฟ ๐—ฟ๐—ฒ๐—ฎ๐—น ๐˜€๐˜‚๐—ฐ๐—ฐ๐—ฒ๐˜€๐˜€.
19
145
4
705
65,511
Gabriele Costantini retweeted
Google just dropped a new LLM! You can run it locally on just 0.5 GB RAM. Let's fine-tune this on our own data (100% locally):
183
1,138
73
14,113
1,957,508