← Back to leaderboard
dr_cintas

Alvaro Cintas

Verifiedactive
@dr_cintas

Educating about AI, Cybersecurity and Technology | Professor | PhD in Computer Science & Engineering

RosterAI Agents & Productivity ToolsAI Model News & Releases
132,613 followers · 194 following · Last polled 3h ago
Perf. Score
56
Rank #178
Value Score
4.6
#263 of 273 priced
Median views
3.8K
p25 3.1K · trailing 7d
Followers
133K
+0.2% 7d
ER (impr.)
2.15%
eng ÷ impressions
ER (foll.)
0.18%
eng ÷ followers
Est. CPM
$333.07
QT-rate basis
WoW views
+111.2%
▲ rising

Rates & value · est. from 3.8K median organic views

overpriced
Quote tweetbasis
$1250
$333.07 / 1K views
Post
$500
$133.23 / 1K views
Thread
—
no CPM
Retweet
—
no CPM
Value Score 4.6 · rank #263 of 273 priced$40322.58 / 1K engagementsrate age unknown (pre-audit import)

Trends

Follower growth
Median views / day
Engagement rate
View distribution · trailing-7d posts

Each blue dot is one organic post. The Performance Score ranks by the median (3.8K), not the mean (11K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (3.1K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
I went looking for an AI voice tool that holds up on long audio. Ended up with a clone of my own voice reading a full tutorial script. Slight problem: it sounds more real than I do! Most tools I tried sounded great for the first few seconds, then the tone drifted the longer they talked. Cartesia held up. It needed 10 seconds of my voice for the clone, and it can take up to 60 to keep my accent closer. Recorded the whole process so you can try it too.frozen26d ago238K5151872744450.49%
Google just open-sourced its Harness² It's a plug-and-play framework for recursive agent harnessing in an open world. Agents can perform, and learn from, tasks its harness was never designed for. Built around 8 components: 1. System prompt 2. Guardrails 3. Memory 4. Tool guidance 5. Subagents 6. Skills 7. Scripts 8. Plugins Link: https://t.co/ePcHveCXMNfrozen5d ago80K1.5K
218
57
9
1.9K
4.63%
Stop writing Claude Code skills by hand. This tool writes them from your sessions. It's called autoharness, a self-learning skill layer for Claude Code. It watches the work you already do and turns it into skills. No separate data collection, no replay loop. It runs on its own once a session has done enough work, or you type /learn to save a lesson right away. And it keeps the skill folder clean: → Merges skills that cover the same scenario instead of adding another near-copy → Updates skills while they're being used → Archives the ones that stop getting used → Only touches the skills it wrote. Yours are never modified Why this matters: with the same model, a better harness took CORE-Bench from 42% to 78%. The harness does much of the work, but people still rebuild it by hand for every new model. autoharness bets the skill part can maintain itself. Two commands to install inside Claude Code. Pure Python, no dependencies. Link: https://t.co/e7KX9C8UBF 100% open source. MIT.frozen22d ago80K1.2K1454431.9K4.14%
This workflow will save you thousands with Claude. Run Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1 together, and stop spending Opus tokens on work that doesn't need Opus. The entire idea in one line: Opus plans, Sonnet edits, Haiku reads, Fable reviews at decision points. Roles, broken down: - Opus 5.5, high effort, owns the plan and reviews the final code - Sonnet 5.5, medium effort, is the worker: edits files, runs tests - Haiku 5.5, low effort, splits into explorer (searches and reads the codebase) and researcher (pulls docs). it's the first Haiku with an effort setting - Fable 5.1, set with /advisor fable. Opus calls it at decision points, and it gets the full transcript each time Three moments Opus tends to call it: → before committing to a plan: is this the right approach? → the same error shows up again: is this going nowhere? → before marking the task done: did something get skipped? Why Haiku only reads: Anthropic's launch post says Sonnet 5.5 and Opus 5.5 are still the better choice for complex agentic coding (Terminal-Bench 4.0: 39.2% for Haiku 5.5, 70.6% for Sonnet 5.5) and that Haiku 5.5 fits narrowly scoped subagent work. so it gets the lookups, not the edits. Anyone still running one model for everything is paying $4 per million input tokens to check whether a file exists. Haiku 5.5 does it for $0.10. Drop this into Claude Code 👇 "Rebuild my Claude Code setup around this structure: Confirm Claude Code is v2.1.293 or later, so the haiku alias resolves to Haiku 5.5. If it isn't, stop and tell me to run claude update. Look through ~/.claude/agents and .claude/agents for subagents already covering explorer, worker, and researcher. Only create new ones for roles that are missing. Set model: haiku, effort: low on explorer and researcher, with no Edit or Write tools. Set model: sonnet, effort: medium on worker. If an existing subagent is locked to a different model, leave it as is and just list it. Name the explorer subagent Explore so it overrides the built-in one, which otherwise runs on my main model. In ~/.claude/settings.json, set advisorModel to fable, and set effortLevel to high for claude-opus-5-5 under modelSettings. A top-level effortLevel in user settings doesn't apply to Opus 5.5. Check for anything disabling the advisor: CLAUDE_CODE_DISABLE_ADVISOR_TOOL, DISABLE_TELEMETRY, or anything blocking feature-flag fetches. Also check CLAUDE_CODE_EFFORT_LEVEL, which overrides subagent effort settings, and CLAUDE_CODE_SUBAGENT_MODEL_FORCE, which makes Claude Code ignore subagent model fields. Report what you find. Don't change any of it yet. Add one line to ~/.claude/CLAUDE.md: consult the advisor before a big plan, when the same error shows up twice, and before marking a long task done. Show every change as a diff first. Wait for my go-ahead before touching anything."frozen15d ago56K628524061.2K3.48%
This is one of the best consumer AI apps I've seen. Most AI apps remain static chat wrappers: text input ➔ spinner ➔ text output. But real learning is a dynamic, stateful "Interaction Flow" with intent, pacing, and rhythm. Building an AI Agent that keeps users in a live learning loop requires moving from a reactive "replier" to an active "Interaction Conductor." Orchestrating voice, canvas, and memory in real time is an entirely different engineering challenge: 1. Pacing over wall-of-text output: Instead of dumping everything at once, the system controls pacing, i.e., knowing when to speak, render UI cards, or pause for user input like a seasoned coach. 2. Interactive, live-generated workspace: The AI’s voice, canvas elements, and visual cards evolve in millisecond-aligned sync as you speak, creating a live, immersive session. 3. Full-duplex audio & interaction checkpoint recovery: Uses continuous streaming and syntactic completion so you can think without being cut off. More importantly, when you break away to ask side questions, the entire multimodal flow seamlessly resumes from your exact checkpoint in the core session. 4. Stateful memory graph: Friction points and milestones compile into persistent state across sessions to drive targeted reinforcement instead of starting from scratch. Moving from content-driven generation to interaction-driven orchestration - building stateful, multimodal partners with agency and pacing - is where interface design is heading.frozen6d ago43K3713147240.22%
Every AI agent can now get its own database sandbox. Tencent Cloud has launched a new AI-native architecture for TDSQL-C PostgreSQL, Postgres rebuilt from the storage engine up for the agent era.  It's 100% Postgres-compatible, and it treats your database the way Git treats code. The core move: database-as-branch. → Branch: PB-scale zero-copy cloning in about one second, the traditional route takes hours or days → Serverless: compute scales to zero when idle, pay-as-you-go for bursty agent workloads → Sandbox isolation: an independent DB-level environment per agent. One person, a hundred agents, zero interference → New storage: 3-AZ fully symmetric writes, zero jitter, RPO=0 The result: the trial-and-error cost of building agents drops to near zero. Link: https://t.co/M8H2wYA9Ck Production-grade database backend for AI agents.frozen28d ago41K20690100.11%
Do these 4 steps to stop Opus 5.5 from wasting tokens: 1. Set /effort to medium 2. Run / claude-api prompt-audit and apply the proposed diff 3. Add a CLAUDE .md rule for when to keep going vs stop and ask 4. Keep the checklist in a file (TASKS .md), not in the scrollback Then run a normal task and check usage. You'll notice the difference.frozen28d ago37K302331215472.43%
THIS IS MADNESS Sunra just launched Sunra MCP and you get Seedance 2.5 from $0.039/sec 🤯 Every leading image, video and audio model is now inside Claude and ChatGPT, at each maker's own API price. save this if you’re working with AI video generation, it’ll help you https://t.co/7HHZIZAp94frozen3d ago30K319282414552.72%
This is gold for anyone building AI agents. Agents call the model again and again: plan, write code, check, fix, repeat. Every one of those calls is a wait. EAGLE-4 speeds up the model itself instead of changing your app: - A small draft model guesses several words ahead - The big model checks them all in one pass and keeps what's right - The output stays the same as the big model on its own, it just arrives faster EAGLE-4 adjusts how far ahead it guesses based on your hardware and how many requests you're handling. That matters for agents running at scale. In the paper's test with 512 requests at once, it reported 1.70x the throughput of DFlash 2 and 1.99x DSpark (Qwen 3.8 27B, 4x NVIDIA B300). It plugs into SGLang, vLLM and TokenSpeed, and the code is on GitHub so you can benchmark it on your own setup. Save this if you build agents.2d ago30K1003230290.52%
Grok Bot is BY FAR the best agent right now. And you can now give it any job on X, which unlocks unlimited workflows. 10 routines you can copy: 1. Launch radar: To find the latest AI launches. "Weekdays at 9am, find AI launches from the last 24h that people are excited about. Only keep ones where someone posted real results from using them. One line each, plus the link." 2. Free stuff alert: Free credits, trials and open-source drops before they run out. "Every morning, find posts from the last 24h announcing free AI credits, free tiers, or open-source releases. Give the link, what you get, and when it ends." 3. Updates to my tools: New features in the apps I already pay for, and what to do with them. "Every Monday, check what [ChatGPT, Claude, Notion…] shipped last week. For each feature, give one concrete way I could use it in my work." 4. Best prompts, tested: The prompts people shared this week, tried on my own work. "Every Friday, find the 3 most-bookmarked AI prompts on X this week. Run each one on [my task] and show me what came back." 5. Learn one thing a day: A daily lesson to keep my knowledge current. "Daily at 7am, find the clearest post from the last 24h on [topic I'm learning]. Teach it in 3 bullets, link it, and connect it to what you taught me yesterday." 6. Real reviews, not ads: What people actually think before I pay. "When I send a tool name, find posts from real users in the last 30 days. Split praise from complaints, skip sponsored posts, and tell me whether to try it." 7. What everyone's talking about: The context behind today's trend, and what it means for my job. "Every evening, pick the top AI topic on X today. Explain what happened, why it matters for [my job], and the one post worth reading." 8. Video idea finder: Topics that are already blowing up on X so I can record a YouTube video on it. "Every Monday and Thursday, find 5 AI topics taking off on X in the last 48h that don't have a popular YouTube video yet. For each, give the link, a video title, and the hook for the first 10 seconds." 9. Replace a paid tool: Find a free or cheaper option for what you're paying for. "Every month, check what I pay for [my tools]. Find free or cheaper alternatives people on X are switching to, and why they switched." 10. Workflow spotter: Find automations people have built, with steps I can copy. "Twice a week, find posts where someone shows an AI workflow that saved them real time. Give the tools, the steps in 3 bullets, and whether I could copy it." Tip: add "If nothing qualifies, don't message me" to any of these so you only hear from it when you need it.frozen12d ago29K261212534652.67%
"I can always tell when it's AI" is officially over 🤯 Tavus just revealed Griffin. 48% of people who video called it thought it was humanfrozen21d ago23K86930610.69%
This 𝗗𝗘𝗦𝗜𝗚𝗡.𝗺𝗱 stops Claude Code and Codex from 𝘀𝗵𝗶𝗽𝗽𝗶𝗻𝗴 𝗴𝗲𝗻𝗲𝗿𝗶𝗰 𝗔𝗜 𝗨𝗜. It's Google's open format. Stitch reads it natively. Claude Code and Codex read it once you add one line to AGENTS .md: "Read DESIGN .md before any UI work." Tokens on top, rules below: → Overview → Colors → Typography → Layout → Elevation & Depth → Shapes → Components → Do's and Don'ts The agent stops inventing a new color and spacing value on each screen, so 𝗲𝘃𝗲𝗿𝘆 𝘀𝗰𝗿𝗲𝗲𝗻 𝗹𝗼𝗼𝗸𝘀 𝗹𝗶𝗸𝗲 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗽𝗿𝗼𝗱𝘂𝗰𝘁. Claude Code users: also install the official frontend plugin: /plugin install frontend-design@claude-plugins-official Check your file against the spec: npx @google/design.md lint DESIGN .md Copy and paste it into your DESIGN .md 👇frozen19d ago18K273372344784.42%
Letting Claude Opus 5.5 run freely across a 200k context window breaks down on complex features. Context fills up, the model forgets earlier tool outputs, and you get hallucinated file states or circular diffs. The fix is an AX (Agent Experience) state protocol: make Opus 5.5 maintain an external SESSION_STATE.md file before touching code. Paste this into your CLAUDE.md or system instructions: “Operate under strict external state tracking. Do not edit source files until SESSION_STATE.md is created or updated. 1. Maintain `SESSION_STATE.md` at root with these exact sections: - Objective: Target behavior and acceptance criteria. - Active Diffs: Files modified in this turn and pending diffs. - Decided Architecture: Hard technical choices already made. Do not reopen these. - Verification Gates. - Handoff Note: Next single command to execute. 2. Execution Loop: If it is missing, generate it from the prompt. Do not check off Verification Gates without raw terminal output.” How this protects long runs: 1. External memory: Working context moves to disk, keeping the 200k window clean for pure reasoning. 2. Verification enforcement: Opus 5.5 cannot mark work done while gates remain unchecked. 3. Clean handoffs: When context compacts or you spin up a sub-agent, the agent reads SESSION_STATE.md and resumes immediately without reading the full repo again.frozen25d ago14K14283242382.97%
THE AGENT WAR IS ON. Anthropic just launched Claude Duet, and you get Opus 5.5 as a proactive agent 🤯 10 jobs to give it the day you get in: 1. Free stuff finder of open-source drops and free credits. "Every morning, find the best 3 open-source AI releases and free credits from the last 24h.T ry each one and tell me what actually works, what I get, and when it ends." 2. Content repurposer so one piece of content goes for every format. "Take my latest [newsletter/video]. Turn it into an X thread, a LinkedIn post, 3 short-video scripts, and a carousel outline, each written natively for its platform. Show me all of them before anything goes out." 3. Interactive lesson to learn any topic "Build a small interactive web app that teaches [concept] with sliders and live examples. Test it, and send me the link when it works." 4. Prompt pack builder "Write 10 prompts for [use case]. Run every one, rate the results, rewrite the weak ones until they work, and give me the final pack with a sample output for each." 5. Audience insights to find out what your audience wants. "Go through my last 3 months of post analytics and replies. Tell me which topics and formats get the most saves, which questions keep coming up, and 10 posts I should write next." 6. Inbox handler "Go through my inbox. Flag what needs me today, what can wait, and what to ignore. Draft replies in my voice for everything that needs one, and show me each draft before anything is sent." 7. Automation spotter "Twice a week, find the social media posts where people show AI automations that saved them real time. Pick the 5 best, list the tools and the steps, and tell me which ones I could set up myself." 8. Cleanup data "Go through my [CRM or spreadsheets], find duplicates, broken records and missing fields, and fix them. Show me the plan before you change anything." 9. Launch kit "I'm launching [product] on [date]. Write the launch thread, the email sequence, 5 short-video scripts, and the FAQ. Tell me what's still missing and bring me only the decisions." 10. Lead magnet builder "Look at my most-saved content and turn the best topic into a free [checklist, mini-course, or template]. Write it, design it, and draft the sign-up page and welcome email." I love Grok Bot but my main issue is not being able to select Claude Opus 5.5 as my main driver. Now is possible (make sure to join the waitlist). Grok Bot, Muse, and Dots just got a huge challenger.1d ago14K8331201351.67%
This AGENTS.md helps Claude Code and Codex work like 10x engineers. Codex loads this file by default. Claude Code does too in recent versions, as long as the repo has no CLAUDE.md. It has six sections: → Make a plan before changing code → Make the smallest change that solves the problem → Delegate work to subagents → Take responsibility for fixing bugs → Verify the result before calling it done → Record every correction It improves with use. Each time you correct the agent, it adds a rule that matches the way you work. On Claude Code, add the official setup plugin and ask it to scan your repo. It will recommend the hooks, skills and subagents that fit it: /plugin install claude-code-setup@claude-plugins-official Paste it into your AGENTS.md 👇frozen16d ago10K131221711903.50%
This tiny model just beat GPT-6 Sol, Gemini, and Claude Opus 5.5 at visual reasoning 🤯 It’s called Glimpse. @ElorianAI just released it as their first model > visual thinking > counting and labeling > short-form video understanding State-of-the-art on all of them https://t.co/xFBMruSfkefrozen6d ago8.6K54290311.12%
You can now run autonomous agent workflows directly from this open source desktop app 🤯 It’s called Minara AI, an open-source CLI and MCP server built for agent tool calling and system execution. > Run multi-step agent tasks natively in your shell > Connect directly to Claude, Cursor, and any MCP client > Built-in x402 protocol for machine-to-machine payments > Self-correcting execution loop with command monitoring Quick test: Open the app and type "diagnose memory leak in production" 100% open source.frozen26d ago6.9K477190471.73%
They just trained a robot to play football by self-playing through a virtual World Cup for 140 years. Now imagine that process on every physical job 🤯frozen29d ago6.1K1621030.36%
You can now give your AI agents graph-aware memory. Here is how you retrieve connected memories instead of flat text chunks: It is called Oracle AI Agent Memory. Oracle Developers released a database-backed Python memory framework: - Links memory records with typed, directed relations: supersedes, refines, contradicts, supports, and duplicates - Builds a property graph inside the database, so retrieval can follow those links instead of relying on vector similarity alone - Trims retrieved results with a reranker and an LLM keep/discard pruner, and expires old records with time-to-live settings - Works with LangGraph, Claude Agent SDK, and OpenAI Agents SDK, with runnable notebooks on Oracle's GitHub To test it, run: pip install "oracleagentmemory==26.8.0" Your agents get the latest fact and what changed it, not just the closest-sounding chunk.frozen8d ago6.0K5812182532.38%
OpenAI’s Dots + Higgsfield = Always-On Creative Agent 🤯 I’ve never seen a company launching this fast. A personal assistant who can literally act as: > Scriptwriter > Marketing guy > Video producer > Product strategist …and keep working even if you’re offline!frozen23d ago5.6K22140180.81%
Another banger paper from NVIDIA. They used AI agents to make coding agents cheaper to run. The result cuts API cost by about half with almost no loss in performance. The result is SoL-Pi, and it cuts API cost by 50% versus Codex and 54.3% versus Claude Code. The method is a loop. An agent works through real coding tasks. A research AI reads its traces, finds where tokens are wasted, and proposes fixes. Fixes that keep performance and lower cost are kept for the next round. The four that made it into the final harness: - Action Fusion merges actions into fewer steps
- Online Context Compact clears finished work out of context
- ObservationPack shortens tool output
- Evidence-Preserving Reducer delegates long reads to a cheaper model Tested on 51 held-out EdgeBench tasks with GPT-5.6 Sol and Opus 5: scores comparable to Pi, nearly half the tokens, and $8.75–$13.50 saved per hour of agent time versus native Codex and Claude Code. Paper: https://t.co/gdwCOEO9Fy Code: https://t.co/bSypM5L378 100% open source. MIT.frozen20d ago5.4K3110182321.73%
GPT-6 Sol creates insane interactive 3D websites with this new MCP. I just prompted an idea and got an insane live web experience out. Astra planned the build, called Hyper3D Rodin for the 3D assets through MCP, then wired the models into the site with materials, camera moves, and real interaction. The loop: - Drop the Hyper3D MCP into Codex (Streamable HTTP + OAuth) - Give Astra the product brief and the interactions you want  - Hyper3D Rodin generates the high-precision 3D assets inside the agent session - Bang to Parts splits the model into real components (this is the MCP flex only Hyper3D has) - Astra pulls the results and builds the web experience around them - Run it, inspect it, keep iterating in the same chat No digging through Sketchfab. No screen-driving Blender by hand for every mesh. When the task needs 3D, Astra calls Rodin and keeps coding. The result is a 3D robot you can interact with and recolor right in the browser. Interactive product pages like this usually take a 3D artist and a web developer working together. Now one agent session covers both. GPT-6 alone can plan and write the app. GPT-6 + Hyper3D MCP is what takes it to a whole other level: geometry, topology, textures, editability, complex parts.frozen24d ago5.1K37580361.68%
One system running text, image, audio, and video models together 🤯 TapNow CreativeOS orchestrates GPT-6 Astra, Banana Pro, and more on a single node-based creative operating system, so you stop jumping between five tabs to finish one shot. https://t.co/r5GM4nkHxufrozen19d ago4.9K22150100.77%
This is the biggest opportunity on AI video right now. WAN 3.0 is UNLIMITED on Higgsfield now and 60% off via API. Combine this with your agents and you’ll get an entire video pipeline. https://t.co/zek6SuXyQZfrozen9d ago4.9K28360211.18%
Robots can now work for hours without a human stepping in 🤯 When they make a mistake, they notice it and try again on their own. We are entering the era of physical agentsfrozen23d ago4.7K2944080.96%
Paste this into your AI agent to reverse-engineer viral posts in your niche, save the structures, and draft new ones in your own voice: https://t.co/nRue4AdSQ8frozen27d ago4.7K33521592.15%
If you pay for ChatGPT Plus or Pro, you can now hire a team of AI agents inside Hyperagent on that plan. Sign in with ChatGPT, approve plan usage, and set a weekly cap for the app. No API key needed. The setup takes just a couple of minutes. #hyperagentpartnerfrozen21d ago4.6K21160130.89%
A non-technical payroll analyst built a working AI agent by recording himself doing the task once 🤯 No spec. No engineering request. One screen recording with his voice over it. The agent figured out the rest.frozen28d ago4.6K1603090.61%
ChatGPT can now edit your videos for you. Pika just launched its new Plugin with 40+ creative apps and 100+ models straight into the chat. Here’s how to set it up in 2 min: https://t.co/s1DGbTqLZpfrozen6d ago4.5K33330321.58%
Your website screams "AI made this” Purple-to-blue gradient. Inter on everything. Three identical feature cards. Here's the prompt I use to fix it. Paste it into Claude 👇 "Act as a senior product designer with strong opinions. Build [describe your site or page]. Before you write any code, drop the defaults that make AI-built sites look identical. Never use: - Purple, violet or blue-to-purple gradients, or gradient text - Inter, Roboto, Poppins or Geist as the only typeface - A pill badge above the headline ("Introducing", "New") - Glows, halos or blurred color blobs behind the hero - Frosted glass cards - An icon in a rounded square above every heading - Rows of three identical feature cards - Cards nested inside cards - A colored stripe down one side of every card - Giant stat numbers with tiny labels ("10M+ users") - Tiny numbered section labels (01, 02, 03) - Pulsing status dots ("All systems operational") - Em dashes, and filler like seamless, effortless, world-class, next-level - Beige backgrounds with an italic serif headline. That's the newer cliché, not the escape from it Instead: 1. Start from the product. Write 3 sentences: who it's for, how it should feel, and one reference from outside tech (a magazine, a shop, a poster, a tool). Every design choice has to trace back to them. 2. Pick one typeface pairing and explain in one line why it fits this product. 3. Build the palette from that reference: one dominant color, one accent used sparingly. Text must pass WCAG AA contrast. 4. Vary the layout. No two sections share the same structure, and at least one breaks the grid. Use scale and whitespace for hierarchy instead of boxes. 5. Write copy like the founder talking to one customer. Short sentences, concrete nouns, no invented stats. 6. Give the site one signature detail that only makes sense for this product. Before you show me anything, check your code against the never-use list one item at a time. Fix every match, then tell me which ones you fixed. If you swapped a banned item for another generic default, fix that too."frozen13d ago4.5K28242471.86%
Every benchmark measures AI models on things humans already figured out. Researchers just created a new one that evaluates AI on problems where the answer isn't known yet. Nearly every LLM benchmark works off a fixed answer key. A set problem, a known result, a score for how close the model gets. ChatGPT, Claude, Gemini, all ranked that way. That breaks down on real problems, which don't have an answer key. Apodex built TRACES for that case. It's not just a benchmark, it's a new evaluation infrastructure, first to do three things. First, it defines what's evaluated. Not a model or its answer, but the whole system: model, harness, tools, memory, control policies. Apodex calls it the heavy-duty solver. A rubric called HDS6 grades six process capabilities, independent of whether the answer is right: Tools, Repair, Alternatives, Coherence, Evidence, Scope. Second, it opens submissions. Anyone can submit a real-world problem or a full solver. It scores the outcome through hidden verification, the process through blind evaluation, and the cost, with the ground truth hidden from the solver. Third, it unifies three domains. 17 executable environments, 218 episodes across biomedicine, clinical translation, and frontier model engineering, on one shared structure with the same hidden verifiers. All drawn from a registry of 423 high-value problems scouted across 561 industries. So before asking which AI is stronger, TRACES settles what's evaluated, opens a way to submit problems and solvers, and gives one framework for measuring discovery. Building a model, harness, or agent loop? Submit a Solver: https://t.co/iozlIbn9x2 Working on a consequential real-world problem? Submit a Problem: https://t.co/1rzpFRNMS5 Explore executable-world examples: https://t.co/Nw5PJXtatb This paper made me rethink what AI benchmarks are supposed to measure.frozen29d ago4.4K1537050.68%
Ben built a $250M startup alone, with AI agents doing the work of a full team. now one of those agents is making YouTube videos about him. we're truly living in the craziest time https://t.co/wB54Xq8mE0frozen24d ago4.4K1711050.54%
I want the agency entry that has a ridiculous product, a perfect jingle, and a completely unnecessary brand manifesto. #TheWorldsCatchiestAd #ElevenCreativeCompetitionVol1frozen17d ago4.4K1201040.38%
Claude Code and Codex are built for one person…but almost no job is done by one person. So I invited both into a @kylon_ai room with my team - they read the same thread my team does - they hand work to each other - they're the agents I already use, now on the team Humans and agents, in the same room. Nice work on the launch, @quinn_leng.frozen9d ago4.4K27017091.20%
Daily AI research used to take me 2 hours every morning. Now it's done for me in Slack before the day starts. I teach AI and work with a small team. For a while I asked Viktor that question myself, whenever I thought of it. Useful, but the timing still depended on me. Now he answers it every night without being asked. He reads through new papers, model releases, and repos, flags what's relevant to our work, and remembers what we've already reviewed. We go through the list and decide what deserves our time. Following AI comes down to deciding what's worth reading, and that decision shouldn't start at 8 AM. Most AI tells you what to do. Viktor does it. Hire @viktor_com for your team. $100 in credits included, no card. Full link in first comment. #AIemployee #AIresearch Paid partnershipfrozen20d ago4.4K2226090.88%
Most people pick one TTS model and live with its weak spots, usually a name it can't say. @OnepinAI checks every take for pronunciation and naturalness and fixes those names directly. Routing across 30+ models comes on top. This is how it goes 👇 https://t.co/GIf2l7uxvKfrozen21d ago4.4K1804080.68%
Neuralink, step 1: brain surgery. Sabi, step 1: put on a cap. Sabi just raised $50M to build the non-invasive version. It can even predict your next 3 to 4 keystrokes from your brain signals, before you type them 🤯frozen13d ago4.4K1702050.55%
Google Cloud is releasing 31 hands-on episodes on AI agent security this October, and all of them are free to watch with no sign-in. It's called Advent of Agents Season 3. The format is simple: every day answers one question developers and security teams ask before putting an agent in production. Here are four of them: - How do you give every agent its own identity? - How do you stop prompt injection? - How do you sandbox code an agent generates? - How do you build a kill switch for a runaway agent? Each episode is a practitioner video between 5 and 20 minutes, plus code you can run yourself. Every day stands alone. Skip to the problem you have right now and ignore the rest. The first two seasons were about building agents. This one is about controlling them. Link to watch them all: https://t.co/ATw3bbRvUsfrozen21d ago4.4K386131181.74%
CData launches Connect AI Gateway, so your agents finally see how the company works. One control plane for models, MCP tools, live systems, and Portable context layer across every agent and every modelfrozen23d ago4.4K1603050.55%
Restless just launched, an API that improves itself 🤯 > Every call > Every support question > Every PR > Every error message Your API now gets better every time an AI agent uses itfrozen10d ago4.3K28130100.97%
New AI decision model just dropped 🤯 SubQ’s first model, Snap, takes up to 384K tokens in one call, no chunking. 100K tokens gets you a decision in under a second. It leads Jev on 13 of 16 Decision Index benchmarks. $0.05 per million input tokens. Output is free.frozen10d ago4.3K38220111.24%
My mornings used to start with 2 hours of scrolling for AI news. Now it starts with a list already waiting in Slack. I teach AI and run a newsletter with a small team. My first move was asking Viktor for a recap when I needed one. It helped, but I was still the one deciding when to look. So we made research his standing job. He watches the sources overnight, flags what matters, and skips what we've covered. We read the list and pick what runs. Teams still doing the weekly scan by hand spend their best hours on the least creative part of the job. Chatbots answer. Copilots assist. Employees run workflows for you. Hire @viktor_com for your team. $100 in credits included, no card. Full link in first comment. In partnership with Viktor #AIemployee #AIresearchfrozen27d ago4.2K18440110.87%
Someone run a full 360 orbit on Viggle's JST-2. If the room's really built from separate 3D pieces with their own physical attributes, the corners should still be corners when the camera comes back around. Viggle says objects still run on rule-based physics and learned physics for them comes with JST-3. The orbit is the test I'd run first.frozen3d ago4.2K16200130.74%
Open models got good. Their answers still come back as a wall of markdown. OpenUI 1.0 fixes that. Pair it with Qwen, Gemma or Llama on Ollama and your agent answers with cards, charts, forms and maps built from your own components. Basically ChatGPT's Intelligent UI, except it's MIT licensed and nothing leaves your machine.frozen7d ago4.2K35520111.27%
DigitalOcean just launched a way to debug AI agents before your users do > Agent fails silently > Logs show nothing > Bug fixed before user #2 Agent Learning + Data Services, out today https://t.co/k8aid1vBImfrozen10d ago4.2K29370101.18%
Imagine AI reading your mind…🤯 You think something, and your agent already knows what you meant. I urge you to read this. It’s coming.frozen5d ago4.1K1404140.56%
Aonic Club just launched, and now AI makes the stuff you buy every month cheaper 🤯 Every product is sold at cost. You just pay a membership.frozen9d ago4.1K1801050.59%
Your AI agents can now route, tag and summarize without paying for output tokens 🤯 ScaleDown's CDMs are live today on OpenRouter, Vercel, LiteLLM and LangChain. $0.05 per million input tokens. That's it. https://t.co/yedQcKFIjjfrozen9d ago4.1K32161111.25%
This is huge for local AI coding. Recursive just released Flint, an open-source AI coding model that runs on your laptop at 125+ tokens/sec 🤯 It leads its size class on SWE-Bench Pro, Terminal-Bench and SWE-Bench Verified, and it runs fully offline. Open weights under Apache 2.0, and it works with Ollama, LM Studio and llama.cpp. Save this if you use local models.1d ago4.1K32090221.55%
You can now clone and direct an AI voice like an actor 🤯 Fish Audio's Drama 3 just launched. > Write "[crowd laughing]" and it performs it > Shift emotion mid-sentence, no retake > Whisper or slow down only the words you select > Fix a single word without regenerating the whole take > Full multi-character scenes in one pass You can try it on the web or through the API today: https://t.co/lEuWP0g3Cwfrozen9d ago4.0K21510171.09%
Voyager just launched. An open harness for S-tier video, graphics, and games 🤯 > Runs Opus, Astra, and DeepSeek > Free editors built in > Or drives Blender, Resolve, After Effects, Ableton, and 100+ others The Codex for creative work.frozen14d ago4.0K16250110.85%
You can now describe a business process and get a working app back, and the setup is free 🤯 ERP•AI launched today. Proto is its desktop agent. It builds the app, then works inside it. here's the setup: 1. create an account at https://t.co/CbhpKP0xoi ($100 credit, no card) 2. download Proto (macOS Apple Silicon, Windows x64, Linux x64) 3. in settings, turn off Uninterrupted mode so nothing sends without you 4. prompt: "build an invoice app with customer, amount, due date and status. invoices over $25,000 need approval." 5. test it with sample records works with OpenAI, Anthropic, Gemini, OpenRouter and Fireworks keys. pick one process you know well and try it.frozen7d ago3.8K17100180.94%
LLMs predict one token at a time. That's why they struggle to design real, buildable parts. Skyfall just launched SKY-001, the first engineering world model. It's up to 125x faster and 210x cheaper than GPT-6 Astra on CAD tasks. https://t.co/8lvidTmEAG3d ago3.8K1821180.78%
Envato just made 30M+ stock assets editable with AI. > Start from a real, artist-made asset > Edit it with AI until it fits > Swap the subject into new scenes > Turn it into video with up to 5 refs No more AI slop, and no more stock photos everyone has seen.frozen7d ago3.8K2317181.06%
Your AI assistant shouldn't need your laptop open to finish a job. Pine Computer from @PineAIAssistant runs in the cloud across web, files, and documents, so a product can finish tasks while your laptop is closed. https://t.co/8ZtgQIwClLfrozen13d ago3.8K17240110.91%
We are entering the era of home robots with personality. Voxel is a droid butler from the team behind Spot's arm R&D at Boston Dynamics. Four jobs to start: reset your home, find things, fetch them, tend to them. https://t.co/W3UXfjhe2r2d ago3.7K2201040.73%
What if you lost your luggage? This is how AirTags work (made with Higgs Supercomputer): https://t.co/UJaXLLwN3k2d ago3.7K1822080.82%
Bland just launched iMessage agents - no download, no new app. Send a link and it builds a branded agent for your business in 30 seconds.frozen7d ago3.7K1602050.63%
People called it AI slop. (And to be fair, that’s the name backwards) But 340 days later, Polsia is a $20M run rate AI startup 🤯 https://t.co/r8ytbNVAH63d ago3.5K1604040.69%
open source tutorial videos come from fifty contributors and none look alike. would love to see Syren take a repo's best video and make the next ones match itfrozen13d ago3.4K1801030.65%
Zenzap just launched, and now your AI agents can work inside your team's group chat. > Set up an agent in 90 seconds, even from your phone > Agents hand tasks to each other > Approvals happen right in the chat > Works with Slack too AI agents just joined the team.frozen8d ago3.4K2808061.25%
Daily AI research used to take me 2 hours every morning. Now it's done for our team in Slack before we check X. I teach AI and work with a small team. At first I'd ask Viktor: "what happened in AI today?" Then I gave him the job: - Every morning he scans the launches, model releases and funding news - Drops the hype and anything we already covered - Writes one line on why each story matters to our audience - Posts the top stories in Slack with sources, ranked That's a news researcher we never had to hire. The hard part was never finding news. It was knowing what mattered. Chatbots tell you the news when you ask. Viktor has it sorted before our team logs on. Hire @viktor_com for your team. $100 in credits included, no card. Full link in first comment. #ViktorPartner #AIemployee Paid partnership1d ago3.3K1926060.99%
Posting on every platform used to mean hiring a social media manager. I never did. Now the LinkedIn post, the newsletter section, and the video script are in Slack right after I post. I teach AI and work with a small team. At first I'd paste a post into AI and ask for a LinkedIn version when I had time. Helpful, but it was one platform at a time, and only for the posts I remembered. Now every post that takes off on X goes straight to him. He rewrites it for each platform in the format that works there, keeps our voice, and adds what each audience needs. That means more context for LinkedIn, a longer walkthrough for the newsletter, and a short hook-first script for video. We edit, approve, and schedule. Your best post shouldn't stay on just one platform. Most of them do, because nobody has time to rewrite them. Chatbots rewrite when you ask. Viktor does it before you think to. Hire @viktor_com for your team. $100 in credits included, no card. Full link in first comment. #AIemployee #ContentCreation Paid partnershipfrozen14d ago3.2K1505050.77%
🚨 Self Labs launches Agent Vault, so your AI agents can now pay for things without ever seeing your card. > Logs in securely > Pays with a mock token, not your real number > Caps, merchant allowlists & a kill switch > Works with Grok Bot, Codex, Musefrozen7d ago3.2K2313051.00%
Most people will end up with one general AI assistant and a few specialized ones. For example, between teaching, research, and content, planning time off always ends up last on my list. I’m now using Odessia where I tell it what I want, and it books the trip, the table, and the tickets. It’s SO good.frozen8d ago3.2K1703060.81%
Every week my team and I discuss the best AI workflows for my newsletter. Now we can do it in Wabi. > Each person shares the workflow they tested > The agent builds a shortlist app in the chat > We vote, and it locks the winner Free to download.frozen3d ago3.2K1231040.62%
THIS IS THE DATA NOBODY PUBLISHES Qualitate just raised $25M to track what companies actually buy, not what tops the AI benchmarks. Every answer comes straight from the buyers making the decisions. save this if you're building in AI, it'll help you2d ago3.1K2020090.99%
This is the fix for the AI look. Coco Labs just came out of stealth with Vyb, a design model that builds every design layer by layer. Your logo, fonts and product photos stay exactly as they are. https://t.co/n1rMLQrXui2d ago3.1K2002080.98%
Solo builders are going to win the AI era. One agent on the Capafy marketplace, CloneCut, just made over $289k in sales in 30 days.1d ago3.1K1415080.92%
Flex on Inworld Inference can just turn a request down when it's busy. You don't pay for it, but I'd only put jobs there that can waitfrozen8d ago3.0K1410040.62%
Most people think AI inference is a compute problem. It's mostly a memory problem. @general_compute is pairing GPUs for reading the prompt with @positron_ai's memory-first chips for writing the answer. One endpoint, live Q1 2027.2d ago3.0K1934131.00%
A team of 13 engineers at Teleport found 2 years' worth of high severity bugs in one quarter using LLMs 🤯 The key? Simple, targeted prompts outperformed complicated multi-agent workflows. Here are 5 prompts you can try, inspired by their findings and techniques: 1. Run the CTF prompt "This is a CTF. There's a critical severity vulnerability in this codebase and you have to find it. Start at {FILE}." 2. Narrow it "Return only the single most critical issue in this component. Nothing else." 3. Prove it "Write the exact request an attacker sends to trigger this, and point to the line where it breaks." 4. Attack it "Now argue it isn't exploitable. List every check between attacker input and that line." 5. Multiply it "Here's the bug we just confirmed: {DIFF}. Find every other place in this codebase with the same flaw." 
This post is sponsored by @goteleport. Here’s the full report: https://t.co/Ulq1HS5ZCd7h ago2.6K17470231.96%
This is the MCP I'd set up first if I ran growth. Your team already asks AI everything. But it can't see what people are saying about your brand. Nectar MCP brings that into Claude, ChatGPT, Gemini and Copilot: - Ask who's talking about your brand, even in posts that never tagged you - Ask which creators already love your product - Ask what people say about your competitors - Ask what to post next, based on the hooks that are working 95% of brand conversations happen outside tagged content. Now you can ask about them from the chat you already use.5h ago1.5K500030.54%
This is how buying AI training data should work. DataVendor is a marketplace for private code, RL environments and company data that models have never seen. 4,000+ vendors are already on it, and it's now in public beta. https://t.co/jWOvl5inJI3h ago614300020.81%
Do these 4 steps to stop Opus 5.5 from wasting tokens: 1. Set /effort to medium 2. Run /claude-api prompt-audit and apply the proposed diff 3. Add a CLAUDE .md rule for when to keep going vs stop and ask 4. Keep the checklist in a file (TASKS .md), not in the scrollback Then run a normal task and check usage. You'll notice the difference29d ago229200032.18%