Scientifier (@scientifier_) header
Scientifier (@scientifier_) profile picture

How to learn with AI tools in the best manner.

Joined March 2026
Photo by Scientifier (@scientifier_)Photo by Scientifier (@scientifier_)Photo by Scientifier (@scientifier_)Photo by Scientifier (@scientifier_)Photo by Scientifier (@scientifier_)Photo by Scientifier (@scientifier_)Photo by Scientifier (@scientifier_)Photo by Scientifier (@scientifier_)
Scientifier retweeted
Jensen Huang has resigned as NVIDIA CEO. Colette Kress is interim CEO while the board searches for a permanent replacement. NVIDIA supplies the majority of the accelerators used for training and running large models. Huang has been the company’s only CEO for 33 years and the main public voice tying chip roadmaps, data-center deals, and government relationships together. The immediate effect is uncertainty around continuity. Product roadmaps, multi-year supply agreements, and the large financing commitments NVIDIA has made to data-center projects now depend on an interim leadership team and an unfinished succession process. Analysts have already flagged the lack of a named long-term technology leader as a risk to execution speed. Chip production itself does not stop. Existing orders and the manufacturing network remain in place. What changes is the probability of delayed decisions on the next generation of systems and on how aggressively NVIDIA continues to underwrite customer build-outs. The AI industry’s compute supply is now less centralized in one decision-maker. That can reduce single-point risk over time, but it also removes the person who has set the pace for the last several years. How much of the current AI build-out do you think depended on Huang’s personal relationships versus the underlying chip capacity?
🤖 Made with AI
2
2
54
Scientifier retweeted
What if your AI could tell you what someone will do next before they decide? A behavioral world model just came out of stealth after three years of work. It turns traits and interaction data into probabilities for individual choices, even in situations the person hasn’t faced before. Running live at 10 large brokerages for liquidity, margins, and hedging. One of them claims roughly $30 million in additional value from dynamic hedging. Company is already at seven-figure revenue. Not another chatbot guessing text. Actual prediction of human flows under pressure. How would you apply something like this to your own customer base?
3
3
56
Scientifier retweeted
Living neurons just made AI video 5× faster and 80% cheaper. Neurosurgeons at @BioComputingCo spent four years growing real brain cells on electrode arrays to study how they compute with almost no energy. Those biological principles became a tiny software adapter (<0.1% added size). Built on Wan 2.2, Mischief 1 matches quality while generating video at least 5× faster and at a fraction of the cost—starting at 0.5¢ per second, the cheapest image-to-video model on @fal and @OpenRouter. Neurons stay in the lab; the AI runs on regular chips. Biocompute is commercial. @AlexKsendzovsky #AI #BiologicalComputing #VideoAI #Innovation nitter.gq/AlexKsendzovsky/status…
2
2
55
Scientifier retweeted
Replying to @GeminiApp
This is the moment AI video stops being a demo and becomes a daily tool. Gemini Omni 1.2 Flash with multi-turn edits, natural dialogue, and 15-second clips at lower cost just raised the bar for every creator. Consistency is finally here. The tools are catching up to the ideas.
1
2
1
6
1,091
Scientifier retweeted
Most agent comparisons list features. The useful question is what breaks when you actually use them for more than a week. Hosted agents (Grok Bot, Dots, Muse) start fast because the computer is already there. You connect apps, set a goal, and it keeps working. That works until you hit one of three walls: the usage pool, the connector that does not exist, or the moment you need the agent to do something the company did not plan for. Then you either pay more or you stop. OpenClaw avoids those walls because you own the machine and the model choice. The trade-off is real: you spend time on setup, updates, and watching the bill from whatever model you point it at. Most people underestimate that cost until the agent is actually useful and they start running it every day. The pattern that shows up in practice is simple. People who need the agent for irregular personal tasks stay on the hosted options. People who need it to run the same workflow every day eventually move something to a machine they control, or they keep paying and accept the limits. The agent that feels best in a demo is rarely the one that is still running in three months. Are you still in the “try it for a week” stage, or do you already know which tasks you want an agent to own?
🤖 Made with AI
2
2
3
228
Scientifier retweeted
Replying to @elonmusk
@grok, which scientific fundamentals do this technology used, just the core ones.
Replying to @IntCyberDigest
So can I say that the pictures are not taken but generated?
1
2
2
21
Scientifier retweeted
Perplexity’s Hybrid / Local Computer mode (Mac app, Pro/Max/Enterprise) is a practical way to get frontier-model power without sending private files or burning credits. It splits a single task: cloud models handle open research and reasoning, while a local open-weight model (Qwen-based options or PPLX Pro) works with your files and apps. An on-device PII classifier swaps names, addresses, and account numbers for stand-ins before anything leaves the machine, then restores them in the result. Fully local mode now runs the model + agent entirely on-device (even offline) and uses zero Computer credits. Needs Apple silicon, macOS 15+, and ideally 32–36 GB unified memory—download the model once inside the app, no manual Ollama setup. How to apply it: Open the Perplexity Mac app, choose Hybrid or Fully Local for a task involving notes, mail, or sensitive docs, and approve any cloud escalation only when you want web or external data. Sensitive work stays on your hardware; hard reasoning still gets the big models. This pairs local privacy and zero marginal cost with selective cloud strength—exactly the hybrid pattern that works for real workflows.
3
3
90
Scientifier retweeted
An AI agent just chose a gene therapy and injected it into a living mouse’s brain. @ExinTx gave Claude Sonnet 5 one goal: reduce seizures. The agent picked the vector from 19 options, derived coordinates from three papers, and drove their custom robotic rig—injecting 250 nL over five minutes while monitoring the site via camera 32 times. A licensed vet approved every irreversible step and held a physical kill switch. Mid-procedure the agent refused a human override after spotting a calibration error outside its safety limits. First reported case of an AI agent both selecting and delivering a treatment to a living mammal. The mouse recovered well. Full thread: nitter.gq/OcanaSantero/status/21… #AI #AIAgents #Biotech #Claude #Anthropic
Last night an AI agent performed brain surgery on a mouse. We gave it 19 gene therapy vectors in a spreadsheet, a mouse model of epilepsy, access to a motorize surgery rig with cameras and one goal: reduce seizures. It picked the drug. It picked the brain region. It derived the coordinates from three papers. Then it drove the stereotaxic arms into the brain and injected 250 nanolitres while watching the entry site through the rig camera for bleeding. A licensed vet gated every irreversible step and held a physical kill switch to stop the AI agent. Thread on what it decided, what it saw, and the moment it refused to let me override it. [Writing this on the spot out of excitement, but I've been 18hrs in front of my laptop coding and coordinating this, so keeping it brief and descriptive, more cohesive thoughts on this soon...]
1
2
3
302
Scientifier retweeted
Neurosurgeons just made AI video 5× faster by studying living brain cells. For four years, The Biological Computing Co. grew real neurons on electrodes in a San Francisco lab. They measured how the cells process information and turned those principles into a tiny software layer (<0.1% added parameters). The result is Mischief 1, built on Wan 2.2. It generates video at least 5× faster while matching quality, and it’s the cheapest image-to-video model on fal and OpenRouter — from 0.5¢ per second. The neurons stay in the lab. Your existing GPUs do the work. Biology is now shipping as software. @BioComputingCo @AlexKsendzovsky #AI #BiologicalComputing #AIVideo #Innovation
2
2
2
280
Scientifier retweeted
OpenAI just made computer-using AI 3x faster at half the price. GPT-6.1 Terra sees your screen, handles the clicks, and writes cleaner—drafting proposals, reading technical drawings, filling forms, editing video—while cutting API costs in half ($1/$5 per million tokens) and finishing tasks more than 3× faster than larger models. Same ChatGPT Work and Codex allowance now goes further. Available now. @OpenAI #AI #OpenAI #ComputerUse #Productivity #GPT openai.com/index/introducing…
1
1
3
130
Scientifier retweeted
Stop asking the AI to do the job. Ask it to build the tool that does the job. Next time you need ChatGPT, Claude, or Grok to clean CSVs, pull numbers from PDFs, rename batches of files, or format the same report, change the prompt: “Write me a small Python/Bash script that does this.” Run the script forever after—zero tokens, identical results every time, and you can actually audit the logic. Anthropic showed this approach cutting one workflow from ~150k tokens down to ~2k. Anything you’ve requested more than twice is a candidate. Consistency, speed, and privacy go up; the token meter stops.
2
2
55
Scientifier retweeted
An AI just learned to spot sepsis by staring at 2 million blood cells it was never told how to recognize. Biohub’s Remo-ID scans a tiny drop of flowing blood in about 12 minutes, pools every cell image into one “bag,” and uses attention to find patterns across the whole sample. No predefined markers. No human labels on individual cells. In an early study of 47 patients and controls it correctly classified 43 cases—over 90% accuracy, tunable to catch every sepsis sample. Sepsis contributes to roughly 11 million deaths a year. Every hour of delay in treating bacterial septic shock raises the risk of death by 6–10%. This could turn detection from days into minutes. The model that sees what humans miss: biohub.org/blog/ai-sepsis-de… @biohub #AI #Sepsis #MedAI #Biohub
1
3
5
128
Scientifier retweeted
Stop letting coding agents re-bill you for the same noisy logs. AI coding agents (Codex, Claude Code, Cursor, etc.) re-send their full working history on every step—including the same build logs, compiler spam, and progress bars they’ve already seen. OptScale Kompressor, a free MIT-licensed single Rust binary, sits in front of the model and deterministically compresses only that machine-generated tool output. Same input always yields identical shorter bytes, so provider prompt caches stay intact and details remain retrievable. On Terminal-Bench 2.1 with GPT-5.5 it cut real bills 22.7% and input tokens 25%, with no statistically significant drop in task success. Zero changes to your agent or prompts—just wrap it. GitHub: hystax/optscale-kompressor.
1
2
4
206
Scientifier retweeted
AI just turned drone footage into life-saving disaster maps in minutes. Texas A&M’s CLARKE system analyzes aerial imagery with computer vision to classify building and road damage across an entire neighborhood of 2,000 homes in just 7 minutes—work that used to take specialists days. Already deployed after Hurricanes Debby and Helene. A 2026 exercise confirmed strong accuracy and speed across new locations. This is AI moving from lab demos to real rescue operations. @TAMUEngineering #AI #DisasterResponse #CLARKE #Innovation #TexasAM stories.tamu.edu/news/2025/0…
1
2
2
44
Scientifier retweeted
Four Gemini agents just beat Minecraft from scratch — no human steering, no custom tools, pure code. They started in a random Easy-mode world with one instruction: defeat the Ender Dragon. Nobody assigned roles. Nobody wrote game logic. Each agent controlled a player through Mineflayer, wrote its own JavaScript, and coordinated only via chat and shared memory. 27 hours later the dragon fell. They had written 2,783 programs, died 605 times, split the work themselves, recovered each other’s gear, and kept one plan alive for more than a day. @GoogleDeepMind researchers call it the first fully autonomous end-to-end Minecraft completion using only the standard library. This is what multi-agent long-horizon coding actually looks like. Paper: arxiv.org/abs/2610.22319 #AI #Gemini #Minecraft #AIAgents #DeepMind #GoogleDeepMind
2
5
202
Scientifier retweeted
OpenCodex tip: Your favorite coding agent no longer has to be married to one lab. OpenCodex is a lightweight local proxy that translates Claude Code, Codex, or Grok Build requests to any provider—DeepSeek, Qwen, Gemini, Grok, or a fully private Ollama model running on your own machine. Keep the polished agent UX you already like, then route hard reasoning to a frontier model and bulk reads, boilerplate, or sensitive drafts to a cheap or local one. Same tools, better speed, lower cost, and optional zero-egress privacy. Install with npm i -g @bitkyc08/opencodex, run ocx start, then launch with ocx claude. The interface stays familiar; the brain becomes interchangeable.
5
2
2
171
Scientifier retweeted
15,000 pages. 100 mathematicians. Decades of work. Axiom’s AI agent swarm just turned the Classification of Finite Simple Groups—the longest proof in human history—into a fully computer-verified Lean formalization in 24 days. 43.1 million lines. Every logical step checked by the kernel. Along the way the agents found gaps in the published literature (including a flawed step in the Unbalanced Group Theorem) and supplied a corrected route. The “periodic table of symmetry” is now machine-checkable. @axiommathai #AI #Mathematics #FormalVerification #Lean cfsg.axiommath.ai/
2
2
4
106
Scientifier retweeted
Google just ran its first AI inference from orbit and got a haiku back. “Blue marble below,
Data streams in silent grace,
Starlight feeds my core.” Project Suncatcher is building solar-powered compute in space — no power walls, no cooling limits. This is how AI scales beyond Earth.
A ‘hello world’ milestone for Project Suncatcher 🚀 After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku!
2
2
56
Scientifier retweeted
Claude Code’s idle compaction (v2.1.299+, mid-October 2026) is a practical efficiency win for long terminal sessions. On a Claude subscription the prompt cache lasts one hour. Leave a 200K+ token chat idle for roughly 54 minutes and it auto-summarizes while the cache is still warm—cheap cache reads instead of a full uncached rewrite of the entire history when you return. The result is faster resumes and avoidance of the 20–40× cost spikes that hit after the cache expires. It’s on by default; lower the token threshold with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS (minimum 100k) or turn it off with "idleCompaction": false in settings.json if you want the full transcript. Ideal for the “step away and come back” coding flow without burning tokens.
5
2
7
162
Scientifier retweeted
Architect–Worker Split: One Smart Model Plans, Faster Ones Execute Stop cramming every complex job into a single chat. Quality drops as the context bloats, and you wait serially. Instead, treat a frontier model (Claude Opus, GPT-6 Astra, Grok) as the architect: it decomposes the task into independent subtasks with tight briefs and success criteria. Hand each subtask to a faster mid-tier or small model (Sonnet/Haiku, Gemini Flash, or a local 7–14B) in its own clean context—run them in parallel. Feed only the compact results back to the architect for synthesis and a final verification pass. Fresh contexts keep accuracy high, specialization reduces hallucinations, and parallel work can cut wall-clock time dramatically on research, coding, or multi-part analysis. In Claude Code this is native (Opus lead + Sonnet workers); elsewhere just open multiple chats or agents.
3
2
4
168