Miguel Miranda retweeted
We're cooked.
terminal2shell.com
Miguel Miranda retweeted
i'm a solo dev and i built an infinitely evolving backrooms game and released it on the web.
35,000 players and ~9,600 hours played. these are my metrics under 36 hours of launch.
each blue dot below is a player on the map i recorded a moment ago. there's ~1,200 players playing the game at THIS moment. SIMULTANEOUSLY. let that sink in.
this is incredible. we're entering a new era for sure.
note for video: only shows 200 players in total because my server chunks the shared space into worlds of 200, so 1200 players would be 6 different worlds and you're seeing one of em.
i'm building out a shared space server that can sustain 2000 players in a single world. launching real soon.
Miguel Miranda retweeted
Portainerโs CEO says Docker users should start planning their move to Kubernetes.
Neil Cresswell says he can no longer recommend Docker long term, citing security concerns around autonomous apps and AI agents.
Portainer 3 will focus new features on Kubernetes. Docker, Swarm and Podman will keep receiving bug and security fixes.
The separate Community Edition is also ending, with CE 2.45 as its final release.
Portainer-D2K offers a migration path that lets users keep familiar Docker tooling on Kubernetes.
Full story: portainer.io/blog/portainer-โฆ
Miguel Miranda retweeted
STOP BURNING OPUS ON F*CKING GREP
Opus 5.5 is your most expensive brain. Stop making it read 2,000-line files and 37 grep hits.
Claude Code subagents fix this.
Opus 5.5 plans the job, then hands off the dirty work.
Haiku 5.5 subagents search the codebase in their own context windows.
Sonnet 5.5 subagents patch and double-check.
A PostToolUse hook runs your tests after every edit.
Only short summaries come back to Opus. The raw sh*t never touches the main context.
You set it up like this:
/agents
The full tree
> Opus 5.5 main session plans + signs off
> .claude/agents/explorer.md โ model: haiku ยท Read, Grep, Glob
> .claude/agents/fixer.md โ model: sonnet ยท Read, Edit, Bash
> .claude/agents/checker.md โ model: sonnet ยท cross-checks every patch
> settings.json โ PostToolUse hook runs tests after each edit
> every subagent gets its own context window
> Opus reads summaries only, never raw output
Paste the tree + this prompt into Claude Code โ
Set up my Claude Code project around this tree:
Create 3 subagents in .claude/agents/:
explorer (model: haiku, tools: Read, Grep, Glob) searches the codebase and returns a summary under 200 words.
fixer (model: sonnet, tools: Read, Edit, Bash) makes the smallest patch that fixes the issue.
checker (model: sonnet, tools: Read, Bash) reviews every patch against its callers and reports regressions.
Add a PostToolUse hook in .claude/settings.json that runs my test command after every Edit/Write.
Main session stays on Opus 5.5: plan, delegate, read summaries only, sign off.
Show me every file as a diff before writing it.
save it, it'll come in handy.
Miguel Miranda retweeted
โผ๏ธ Storage maker QNAP says customers who ran an old Plex media server open to the internet should assume their NAS has been hacked.
Attackers chain four Plex flaws, dubbed Plex2Shell, to run their own code on the server without logging in. QNAP says hijacked servers have been used to mine crypto and run proxy software.
Some QNAP users' Plex servers already appear compromised. CISA lists the flaws as actively exploited.
Plex fixed them in August with version 1.43.3.10861. Updating now won't remove what attackers installed, so QNAP says to reinstall and reset NAS and Plex passwords.
Miguel Miranda retweeted
Turn a single image into a 3D world you can explore with Lyra 2.0.
By combining video generation with 3D reconstruction, our researchers create environments along a camera path you choose. Lyra draws on earlier views and learns to correct accumulating visual errors, helping the world stay consistent as you explore and return to familiar places.
Paper, code, and demos: nv-tlabs.github.io/Project-Lโฆ
Miguel Miranda retweeted
โผ๏ธ Revolut says it cut AI coding costs about 10x by making a self-hosted Chinese model its default in place of Claude Code.
It runs Z ai's GLM on its own Nvidia GPUs for about $0.30 per million tokens.
Miguel Miranda retweeted
Cities, cars, trucks, busses, trains, headlights, break lights, turn signals, water, clouds, rain, bridges, < 500k, no libraries #webgpu
greggman.github.io/gen-city-โฆ
Miguel Miranda retweeted
Harness Engineering 2.0 is here!
We benchmark models all the time but the harness running the model can dramatically improve the result, cost, and execution time too.
And finally, devs just open-sourced Harness Arena, where you can test the same task across multiple harness and model combinations side by side.
You provide one task and then run it across several model x harness configs that you want to test.
Every config receives the same prompt, files, tools, skills, and evaluation checks. Harness Arena then returns:
- total cost
- execution time
- whether the task passed
- files and artifacts produced
If you want to try it yourself, it is implemented within the open-source HarnessRouter project.
Here's the repo: github.com/HarnessRouter/harโฆ
(don't forget to star it โญ )
HarnessRouter is a plug-and-play infrastructure layer that lets you run any harness under a single interface, like:
- Codex
- Hermes
- Claude code
- DeepSeek Harness
- System One, powered by Jev
- And 25 more agent harnesses
In Harness Arena, the cheapest run does not automatically win.
A config must complete the task correctly first. Cost and execution time are compared only among the configurations that pass.
Here's the result from one Harness Arena test:
- Hermes + GPT-5.2: 5/5 successful runs using 0.47 credits
- Claude Code + Opus 4.8: 3/5 successful runs using 223 credits
That is around 474x difference in credits, while the more expensive configuration also failed two runs.
You can keep the model fixed to measure how much the harness changes the result. Or vary both to find the best harness and model pairing for that task.
This helps answer practical questions such as:
- Should this task run on Codex or Claude?
- Can a cheaper config still pass every check?
- Which setup is best to handle this workload?
Once you find the best config, you can run future tasks through it using the same Unified Harness Protocol (UHP) interface for sessions, streaming, files, artifacts, cancellation, and failures.
So you are no longer choosing an agent stack based only on a model leaderboard.
Instead, you can benchmark the complete runtime on the task your product actually needs to solve.
To dive deeper, my co-founder wrote an article explaining why model routing is different from harness routing, how UHP works, and what supporting multiple agent runtimes involves.
It also covers the local setup, a working request, and how sessions and files move through the system.
Read it below.
Miguel Miranda retweeted
Four weeks of wall squats reduced systolic blood pressure by nearly 10 points in adults with high blood pressure.
Researchers then compared continuing three sessions weekly with reducing to two, one, or none for another month. Systolic pressure remained roughly 7 mmHg below starting levels in the once- and twice-weekly groups. Continuing three sessions weekly produced the greatest reduction of about 14 mmHg. In those who stopped completely, blood pressure moved back toward baseline.
This fits with a previous analysis of randomized trials that identified isometric exercise, which involves holding a muscle contraction, as an effective way to lower resting blood pressure.
What interests me is the practical question - how much of an exercise benefit can we preserve when we temporarily have less time? This study suggests that reducing the frequency can retain part of the benefit for up to a month.
Miguel Miranda retweeted
๐จ NVIDIA JUST OPEN-SOURCED A SEMANTIC LAYER FOR AI AGENTS
Itโs called Auto Ontology, and it tackles a problem that better SQL alone wonโt fix.
An agent can query the right database and still return the wrong answer.
Why?
Because โrevenueโ might mean something different in every system.
Auto Ontology is a Python framework that builds shared context over enterprise metadata:
โ Entities and business metrics
โ Relationships between data
โ Certified definitions and lineage
โ Provenance behind answers
It can generate ontologies from metadata, connect existing semantic models, and reconcile conflicting definitions.
For structured data, it also supports natural-language-to-SQL with validation against the ontology.
Agents can access that context through MCP, CLI, or API.
The useful shift: giving agents access to business meaning alongside the data.
Open source under Apache 2.0.
Repo link in ๐งต โ
Miguel Miranda retweeted
BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list.
Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic and OpenAI. Engineering lead time improved 70%.
Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything.
Frontier models become the exception, open weights become the default. That's a pricing problem for every lab heading to IPO.
h/t @OliverMolander
Would you move your team off frontier APIs if the bill dropped 90%?
Miguel Miranda retweeted
10 GitHub repos every AI developer should know before putting an AI app into production.
Building an AI application is one thing. Making sure it stays reliable, accurate, and cost-efficient is another.
These open-source tools help you understand what's happening behind the scenes:
1. Langfuse
Track LLM calls, prompts, performance, and costs.
github.com/langfuse/langfuse
2. Arize Phoenix
Trace AI workflows and debug retrieval problems.
github.com/Arize-ai/phoenix
3. Promptfoo
Test prompts, compare model outputs, and catch regressions.
github.com/promptfoo/promptfโฆ
4. DeepEval
Evaluate LLM responses, hallucinations, and RAG performance.
github.com/confident-ai/deepโฆ
5. TruLens
Check whether your RAG application retrieves relevant information and generates grounded answers.
github.com/truera/trulens
6. LiteLLM
Connect multiple LLM providers through one API with routing and spend tracking.
github.com/BerriAI/litellm
7. Helicone
Monitor LLM requests, latency, usage, and spending.
github.com/Helicone/helicone
8. MLflow
Track experiments and evaluate machine learning and LLM applications.
github.com/mlflow/mlflow
9. OpenLIT
Monitor LLM traces, metrics, and GPU usage with OpenTelemetry.
github.com/openlit/openlit
10. Oxylabs Web API MCP
Bring live web search and extracted data into AI workflows.
github.com/oxylabs/web-api-mโฆ
A working AI app is just the beginning.
You also need to know when it gives poor answers, where it spends money, and what needs fixing.
That's what separates a quick AI prototype from a reliable product.
Save these repos before your next AI deployment. ๐
Miguel Miranda retweeted
NVIDIA Open-Sources Auto Ontology, a Semantic Layer for AI Agents
Your boss asks the AI agent: "What was revenue last quarter?"
It answers in 3 seconds. Clean number. Nice chart.
But it used the wrong "revenue" column. ๐ก๐ผ๐ฏ๐ผ๐ฑ๐ ๐ป๐ผ๐๐ถ๐ฐ๐ฒ๐.
That's the problem NVIDIA's new open-source library, Auto Ontology, goes after. Its README puts it plainly:
"โฆ๐ข ๐ณ๐ฆ๐ท๐ฆ๐ฏ๐ถ๐ฆ ๐ฎ๐ฆ๐ต๐ณ๐ช๐ค ๐ข๐ฏ๐ด๐ธ๐ฆ๐ณ๐ฆ๐ฅ ๐ง๐ข๐ด๐ต๐ฆ๐ณ ๐ช๐ด ๐ธ๐ฐ๐ณ๐ต๐ฉ ๐ฏ๐ฐ๐ต๐ฉ๐ช๐ฏ๐จ ๐ช๐ง ๐ต๐ฉ๐ฆ ๐ฏ๐ถ๐ฎ๐ฃ๐ฆ๐ณ ๐ช๐ด ๐ธ๐ณ๐ฐ๐ฏ๐จ."
๐ช๐ต๐ฎ๐ ๐ถ๐ ๐ถ๐:
A "context layer" that sits between an AI agent and your database. It tells the agent where the data lives, what each field means, and whether to use it.
๐๐ผ๐ ๐ถ๐ ๐๐ผ๐ฟ๐ธ๐, ๐๐๐ฒ๐ฝ ๐ฏ๐ ๐๐๐ฒ๐ฝ:
โข ๐๐ถ๐๐ฐ๐ผ๐๐ฒ๐ฟ: it reads your database metadata and drafts an ontology, a map of business terms like "active customer"
โข ๐๐ฟ๐ผ๐๐ป๐ฑ: agents use those terms instead of guessing from raw table names
โข ๐ฉ๐ฎ๐น๐ถ๐ฑ๐ฎ๐๐ฒ: it checks generated SQL against certified definitions
โข ๐๐
๐ฝ๐ผ๐๐ฒ: agents reach it through MCP, a CLI or an API
โข ๐๐ฒ๐ฎ๐ฟ๐ป: certified corrections go back into the map
๐๐ผ๐ฟ ๐๐๐ฒ๐ฟ๐:
Ask a question, get ๐๐ต๐ฒ ๐ฎ๐ป๐๐๐ฒ๐ฟ, ๐๐ต๐ฒ ๐ฆ๐ค๐ ๐ถ๐ ๐ฟ๐ฎ๐ป, ๐ฎ๐ป๐ฑ ๐๐ต๐ฒ ๐ฟ๐ผ๐๐. You can check the work, not just trust it.
๐๐ผ๐ฟ ๐ฏ๐๐ถ๐น๐ฑ๐ฒ๐ฟ๐:
โข Apache 2.0, runs with Docker Compose or Helm
โข Works with Databricks, Postgres, Snowflake and DuckDB
โข Defaults to NVIDIA's Nemotron models, but ๐ฎ๐ป๐ ๐ข๐ฝ๐ฒ๐ป๐๐-๐ฐ๐ผ๐บ๐ฝ๐ฎ๐๐ถ๐ฏ๐น๐ฒ ๐ฒ๐ป๐ฑ๐ฝ๐ผ๐ถ๐ป๐ ๐๐ผ๐ฟ๐ธ๐
โข The MCP server has ๐ต ๐ฟ๐ฒ๐ฎ๐ฑ-๐ผ๐ป๐น๐ ๐๐ผ๐ผ๐น๐. There is deliberately no tool to browse raw tables, so agents must go through the business terms
๐๐ ๐ฏ๐๐ถ๐น๐ฑ๐ ๐ผ๐ป ๐ฒ๐ฎ๐ฟ๐น๐ถ๐ฒ๐ฟ ๐๐ผ๐ฟ๐ธ:
โข Cube and dbt MetricFlow: semantic layers you write by hand in code
โข Wren AI: an open semantic layer for AI agents, with a context layer built partly by agents
โข Snowflake's Semantic View Autopilot: also drafts semantic models with AI, but only for Snowflake
What's different here: ๐ผ๐ฝ๐ฒ๐ป ๐๐ผ๐๐ฟ๐ฐ๐ฒ, ๐๐ผ๐ฟ๐ธ๐ ๐ฎ๐ฐ๐ฟ๐ผ๐๐ ๐ณ๐ผ๐๐ฟ ๐ฑ๐ฎ๐๐ฎ๐ฏ๐ฎ๐๐ฒ๐, ๐ฎ๐ป๐ฑ ๐ฏ๐๐ถ๐น๐ ๐ณ๐ผ๐ฟ ๐ฎ๐ด๐ฒ๐ป๐๐ ๐๐ต๐ฟ๐ผ๐๐ด๐ต ๐ ๐๐ฃ.
๐ง๐ต๐ฒ ๐ต๐ผ๐ป๐ฒ๐๐ ๐ฝ๐ฎ๐ฟ๐:
โข ๐ง๐ต๐ฒ "๐ต๐ฌ% ๐ณ๐ฎ๐๐๐ฒ๐ฟ" ๐ฐ๐น๐ฎ๐ถ๐บ ๐ถ๐ ๐ก๐ฉ๐๐๐๐'๐ ๐ผ๐๐ป ๐๐ฒ๐๐๐ถ๐ป๐ด, with no baseline published
โข ๐ก๐ผ ๐ฎ๐ฐ๐ฐ๐๐ฟ๐ฎ๐ฐ๐ ๐๐ฐ๐ผ๐ฟ๐ฒ๐ ๐ฎ๐ฟ๐ฒ ๐ฝ๐๐ฏ๐น๐ถ๐๐ต๐ฒ๐ฑ. NVIDIA ships an eval harness (BIRD, 1,534 questions) but no results
โข That harness was tested with ๐๐น๐ฎ๐๐ฑ๐ฒ ๐ข๐ฝ๐๐ ๐ฐ.๐ด, ๐ป๐ผ๐ ๐ก๐ฒ๐บ๐ผ๐๐ฟ๐ผ๐ป
โข ๐ช๐ถ๐๐ต๐ผ๐๐ ๐๐ฎ๐๐ต๐ถ๐๐ผ๐ฟ๐ฝ ๐ฉ๐ฎ๐๐น๐, ๐ฑ๐ฎ๐๐ฎ๐ฏ๐ฎ๐๐ฒ ๐ฝ๐ฎ๐๐๐๐ผ๐ฟ๐ฑ๐ ๐ฎ๐ฟ๐ฒ ๐๐๐ผ๐ฟ๐ฒ๐ฑ ๐๐ป๐ฒ๐ป๐ฐ๐ฟ๐๐ฝ๐๐ฒ๐ฑ in Postgres (per the README)
โข It's young: ๐ฎ๐ฏ๐ผ๐๐ ๐ณ๐ฑ๐ฌ ๐๐ถ๐๐๐๐ฏ ๐๐๐ฎ๐ฟ๐, and NVIDIA isn't accepting outside contributions
๐ง๐ต๐ฒ ๐ฏ๐ถ๐ด๐ด๐ฒ๐ฟ ๐ฝ๐ผ๐ถ๐ป๐: the hard part of AI analytics isn't writing SQL. ๐๐'๐ ๐ฎ๐ด๐ฟ๐ฒ๐ฒ๐ถ๐ป๐ด ๐ผ๐ป ๐๐ต๐ฎ๐ ๐๐ต๐ฒ ๐๐ผ๐ฟ๐ฑ๐ ๐บ๐ฒ๐ฎ๐ป.
GitHub: github.com/NVIDIA/auto-ontolโฆ
Links in reply ๐
Miguel Miranda retweeted
NuGet vulnerable with Microsoft.Data.SqlClient 7.1.1 depending on System.Security.Cryptography.Pkcs 9.0.18 stackoverflow.com/q/80013443โฆ
Miguel Miranda retweeted
80,000 lines of code. 92,000 trees. Real-time ray tracing.
Two weeks of development using Claude Code to build a complete off-road racing game from scratch on a custom C++/Vulkan engine.
Source:
github.com/markaren/threepp
Miguel Miranda retweeted
ในใใใฎๅพใใซๅใใใฆใ็ตตใๅใใใใซใใฆใฟใใ
3Dๅใใฆใใใใใงใฏใชใใ1ๆใฎ็ตตใใ360ยฐๅ่ปขใฎใๅ็ปใใไฝใใๅพใใซๅใใใฆใใฌใผใ ใๅๅพใใใฆใ็ซไฝใฃใฝใ่ฆใใฆใใพใใ
#MiniMaxH3 + ORB360 LoRA
Miguel Miranda retweeted
โผ๏ธ Ubuntu and Red Hat still have no fix for a Linux kernel flaw that Anthropic says its unreleased AI model turned into a remote root exploit.
The bug, CVE-2026-98432, is a use-after-free in the IPv4 TCP handshake, rated 9.8 out of 10.
The exploit needs no credentials and targets servers with a multipath TCP listener, which Go 1.24 and later open by default.
The fix reached stable kernels on 11 October and Debian has shipped it. Ubuntu still marks every release from 16.04 to 26.10 vulnerable.
Red Hat now rates it Critical but has no patched kernel for RHEL 8, 9 or 10.
Anthropic has not released the exploit. By its count, it takes 48 minutes to leak a kernel pointer and five more to run commands as root.
Miguel Miranda retweeted
What a disaster atlasum.net/blog/cve-2026-98โฆ
Miguel Miranda retweeted
Hard to overstate how big of a deal this is.
tl;dr โ ANY unpatched Linux server with even a single TCP port exposed to the internet (say for HTTP or SSH) and MPTCP active (default for many Go web servers) is vulnerable to complete takeover.
What a disaster atlasum.net/blog/cve-2026-98โฆ