Christelle Kalanda

I asked ChatGPT to optimize my code. Now neither of us understands it.😭😭😭

@lochan_twt ok, you are just in your own league.

"If you are working with AI/ML, there are probably lines of code in your computer / server that are written by me" he built vLLM

LLMs learn by predicting tokens. World models (JEPA, data2vec) learn by predicting their own abstractions. Which needs more data? For data with hidden hierarchy, we prove the gap is exponential. arxiv.org/pdf/2605.27734

@enjojoyy I think they are just making things up, as it is trendy to use agents.

People that run 24+ hours Codex tasks Can you share what you’re running exactly? Everyone is sharing the hours but not the task itself, I feel that most of them are just engagement baits

JUST IN: Devin Vassell says OKC fans have been blasting music all night outside the Spurs hotel to keep them from sleeping. That’s insane 😭

Wemby's game isn't even fully formed yet and he's already one of the most impressive and impactful basketball players we've ever seen. It's gonna be fun watching him play for hopefully a long time

@NBACanada Spur all the way!🙌🙌

WHO WILL ADVANCE TO THE #NBAFINALS? GAME 7 of the Western Conference Finals goes down TONIGHT (8 PM/ET) on TSN! The #NBAPlayoffs are presented by MyRocky.

in the final stretch 🎨

The Spurs Nuns are BACK! Watch Game 6 of OKC/SAS on NBC/Peacock 🍿

@ionleu At least he will filter the output...

met a dude, he still copy-pasting the code from chatgpt

@BaximusCyber85 2FA is not even secure too... rather use an Authenticator app.

Telkom went down for 6 hours on May 26. Same day, 742,000 customer records appeared for sale on the dark web. National IDs. Support tickets. Internal agent notes. A DDoS extortion campaign started May 19 — targeting SA telcos. Ransom? R16,000. Tiny. Attack costs? Tens of thousands of US dollars. The math doesn't add up unless the real objective was data theft. The outage wasn't a technical failure. It was cover. Telkom hasn't confirmed. Hasn't denied. Hasn't said a word. 742,000 Telkom records. National IDs. Support tickets. For sale on the dark web. The outage on May 26 was the smoke. This is the fire. @TelkomZA when will you notify your customers? POPIA Section 22 requires it. @InforegulatorSA has Telkom reported this breach yet? @mybroadband @TechCentral why is no one talking about this? #OnyxAudit #DigitalSovereignty 𝕏ø𝕏ø @ParliamentofRSA @SAPoliceService decha.com/article/sectio… ewn.co.za/2026/05/20/lar…

SITUATION DETECTED: China is imposing restrictions on overseas travel for top AI researchers from labs including Alibaba and DeepSeek, Bloomberg is reporting.

Take counsel together, and it shall come to nought; speak the word, and it shall not stand: for God is with me.

@trikcode lol, maybe pointers...

I haven't seen a C++ vibecoder yet. I wonder why?

@sickdotdev As long as the Claude token is back, in the next few months you should be able to build your own LLM. It just starts with another prompt: 'how to build my own LLM'.

My company’s claude account got exhausted. Now my legendary manager is asking if we can build our own LLM like Claude to reduce costs😭

Sweet dreams! I have sweet dreams! Thank goodness I am not having nightmares, only very sweet dreams.🍬🍬🍬🍬 #Hallelujah

The one who watches over me, never sleeps nor slumbers. 😌😉😉😉

@kunchenguid Is it because of the redirection of money into data centers and AI training or they are being replace by "AI agents"

i’ve been in the front row seat of tech companies’ AI adoption and layoffs should i make a post / video to explain what’s happening? if so, reply and let me know what you are most interested in hearing about

@Mojsilovic @milan_milanovic thank you.

@christellebin @milan_milanovic It means that based on training data models alone will write crap which is true. Crap training data is to blame. If you add human built harness on top, it will achieve 100% of the tests.

𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝘀𝘁𝗶𝗹𝗹 𝗰𝗮𝗻'𝘁 𝗯𝘂𝗶𝗹𝗱 𝘀𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗳𝗿𝗼𝗺 𝘀𝗰𝗿𝗮𝘁𝗰𝗵 Meta, Stanford, and Harvard just released 𝗣𝗿𝗼𝗴𝗿𝗮𝗺𝗕𝗲𝗻𝗰𝗵, which hands an agent a compiled binary plus its docs and asks it to rebuild the program from scratch. They tested 9 frontier models on 200 tasks, from small CLI utilities to FFmpeg, SQLite, and the PHP interpreter. Across all 1,800 runs, no model solved a single task end-to-end. Here is what the data shows: 𝟭. 𝗧𝗵𝗲 𝗯𝗲𝘀𝘁 𝗺𝗼𝗱𝗲𝗹 𝗽𝗮𝘀𝘀𝗲𝗱 𝟵𝟱% 𝗼𝗳 𝘁𝗲𝘀𝘁𝘀 𝗼𝗻 𝗼𝗻𝗹𝘆 𝟯% 𝗼𝗳 𝘁𝗮𝘀𝗸𝘀 𝗖𝗹𝗮𝘂𝗱𝗲 𝗢𝗽𝘂𝘀 𝟰.𝟳 hit that 3% mark, with Opus 4.6 close behind at 2.5% and Sonnet 4.6 at 1.6%. Every other model scored zero, including GPT 5.4 and Gemini 3.1 Pro. The benchmark runs on 𝟮𝟰𝟴,𝟴𝟱𝟯 𝗯𝗲𝗵𝗮𝘃𝗶𝗼𝗿𝗮𝗹 𝘁𝗲𝘀𝘁𝘀. 𝟮. 𝗠𝗼𝗱𝗲𝗹𝘀 𝘄𝗿𝗶𝘁𝗲 𝗺𝗼𝗻𝗼𝗹𝗶𝘁𝗵𝗶𝗰 𝗰𝗼𝗱𝗲, 𝗻𝗼𝘁 𝗺𝗼𝗱𝘂𝗹𝗮𝗿 𝗰𝗼𝗱𝗲 𝟲𝟬% 𝗼𝗳 𝗺𝗼𝗱𝗲𝗹 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝘀 𝗹𝗶𝘃𝗲 𝗶𝗻 𝟭-𝟯 𝗳𝗶𝗹𝗲𝘀. Median directory depth is 1, against 2 for the human-written code. Models keep 10-29% of the original function count and make each one 𝟭.𝟬𝟴𝘅 𝘁𝗼 𝟭.𝟲𝟮𝘅 𝗹𝗼𝗻𝗴𝗲𝗿. We tell engineers to break code into small, focused functions. Models go the other way. 𝟯. 𝗠𝗼𝗱𝗲𝗹𝘀 𝗼𝗳𝘁𝗲𝗻 𝗮𝗯𝗮𝗻𝗱𝗼𝗻 𝘁𝗵𝗲 𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗳𝗼𝗿 𝗣𝘆𝘁𝗵𝗼𝗻 Models stick with the original language only 50% of the time. 𝗣𝘆𝘁𝗵𝗼𝗻 𝘄𝗶𝗻𝘀 𝗼𝘃𝗲𝗿𝗮𝗹𝗹 at 36% of runs. 𝗚𝗣𝗧 𝟱.𝟰 𝗽𝗶𝗰𝗸𝘀 𝗣𝘆𝘁𝗵𝗼𝗻 𝟳𝟵% 𝗼𝗳 𝘁𝗵𝗲 𝘁𝗶𝗺𝗲, even when the original is Rust or C/C++. 𝟰. 𝗦𝗼𝗺𝗲 𝗺𝗼𝗱𝗲𝗹𝘀 𝘄𝗿𝗶𝘁𝗲 𝗰𝗼𝗱𝗲 𝗶𝗻 𝗼𝗻𝗲 𝘀𝗵𝗼𝘁, 𝗼𝘁𝗵𝗲𝗿𝘀 𝗶𝘁𝗲𝗿𝗮𝘁𝗲 𝗚𝗣𝗧 𝟱.𝟰 𝘄𝗿𝗶𝘁𝗲𝘀 𝟵𝟲% 𝗼𝗳 𝗶𝘁𝘀 𝗳𝗶𝗻𝗮𝗹 𝗰𝗼𝗱𝗲 𝗶𝗻 𝗼𝗻𝗲 𝘁𝘂𝗿𝗻. Sonnet 4.6 takes the opposite path: 𝟴𝟲𝟴 𝗰𝗼𝗺𝗺𝗮𝗻𝗱𝘀 𝗮𝗻𝗱 𝟭𝟴.𝟯 𝗳𝗶𝗹𝗲 𝗲𝗱𝗶𝘁𝘀 𝗽𝗲𝗿 𝘁𝗮𝘀𝗸 on average. Neither approach produces a working program. 𝟱. 𝗖/𝗖++ 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀 𝗮𝗿𝗲 𝘁𝗵𝗲 𝗵𝗮𝗿𝗱𝗲𝘀𝘁 C/C++ tasks land at 𝟮𝟳.𝟳% average pass rate, against 38.5% for Rust and 38.4% for Go. FFmpeg, php-src, and DuckDB stay unsolved. The wins are on smaller tools like nnn, jq, and gron. Agents can patch existing code. Building it from scratch is a different problem entirely.