Elon Musk
@elonmusk ·
Our second product is called Blindsight and it will, as the name suggests, enable even those who are completely blind or have never seen to see.
It is ready from an engineering standpoint. Just waiting on regulatory approval.
Back
AI posts
AI posts
People and companies shaping AI. Filter your watched accounts, catch up on recent posts, or find the most liked and viewed.
Compiled by the Toolstash team · Updated
209 watched accounts · 2983 matching posts · Showing up to 3,000 recent stored posts. Counts are snapshots, not live X totals. This page refreshes every minute while visible. Last collected .
Andrew Curran@AndrewCurran_ ·
North Korean state hackers appear to have been behind the Binance attack. The entire event started through a single DevOps engineer's compromised laptop. With the confirmed impact now at $6.3 billion, it is by far the largest crypto heist in history.
- Elon Musk
@elonmusk ·
BABY, YOU FLEW
Thor 雷神 ⚡️@thorwebdev ·
😎
roon@tszzl ·
6 Astra is an old model at this point but inside the codex harness it has a level of AGI-ness that all the other frontier models lack. computer and internet use are seamless and Just Work™️; covers almost anything a person can do on a computer
Shubham Saboo@Saboo_Shubham_ ·
wtf is an agent harness.
An agent is a model plus a harness.
Model reads the context and picks the next action. The harness is all the code that turns that choice into work.
Claude Code, Codex, Antigravity, OpenClaw and Hermes Agent are all harnesses.
8 common parts every agent harness shares:
1. The loop • Sends the context to the model and runs whatever tools it asks for • Feeds the results back in for the next step • Decides when the task is finished
2. Tools and sandbox • Give the model a terminal, files, a browser and your connected apps • Run every task inside its own isolated computer • Keep the damage from a bad command inside the sandbox
3. Subagents • Take side tasks off the main agent and run them in parallel • Start each one with a fresh context window • Send back only a short summary, so the main context stays clean
4. Context management • Decides what the model sees on every turn • Loads skills and files only when the task needs them • Summarizes old history before the context window fills up
5. Memory • Stores progress notes, git history and skills on disk • Lets a new session read what the last one did • Picks the work up where it stopped
6. Guardrails • Hold risky actions like emails, payments and deletes for your approval • Keep real credentials outside the sandbox • Give a prompt injection no credentials to steal
7. Verification • Stops the agent from handing you a summary that only sounds finished • Runs tests, browser checks or a separate grader first • Lets the agent call the task done only after those checks pass
8. Observability • Traces every model call and tool call • Spots the agent repeating the same failed edit • Pushes it to try a different approach
follow @Saboo_Shubham_ for more on Agent Harness.
Joshua Saxe@joshua_saxe ·
The security sensitivity of sites like OpenRouter is insane; imagine the damage you could do by inserting shell commands into people's token streams or just logging and ransoming people's data. Who needs to poison model weights when you can implant yourself within inference providers
dax@thdxr ·
what if you could use your claude subscription in opencode
Daniel San@dani_avila7 ·
Took this photo this morning…
Fall in Michigan is absolutely beautiful
Dean W. Ball@deanwball ·
“It may be the devil, or it may be the lord, but you’re gonna have to serve somebody”
Peter Yang@petergyang ·
I have never seen so many cartoon avatars getting angry at @ForwardEditor for talking about the inevitable.
Gergely Orosz@GergelyOrosz ·
Correct, NVIDIA has a $10K/dev/month limit, a dev a trading firm I talked to (who wants to keep it anonymous) said its $25K and mainly traders who get close
Full: https://newsletter.pragmaticengineer.com/p/whats-the-roi-of-software-engineers
- dax
@thdxr ·
i made and ate two steaks a day for two months
- Elon Musk
@elonmusk ·
Tomorrow is the last day of early voting in Texas.
Vote for @KenPaxtonTX!
kache@yacineMTB ·
There is no amount of money you could pay me to be a DevOps engineer for a crypto exchange
- Daniel San
@dani_avila7 ·
The IDE is not dead… but the way we build software is changing
Today I gave a talk at Latam Build: “The Rise of the ADE” (Agentic Development Environment)
I shared how I’m moving from writing code to orchestrating agents, and what six weeks of working without an IDE taught me
Here are all 23 slides 🧵
- roon
@tszzl ·
rip king
- dax
@thdxr ·
so did all those people buying mac minis for openclaw not use them long enough to hit this
swyx@swyx ·
excited for @Cloudflare's debut as a frontier lab
#NothingButNET gang, assemble!
- kache
@yacineMTB ·
- kache
@yacineMTB ·
Artificial Analysis@ArtificialAnlys ·
Utopai X (based on MiniMax H3) takes #2 on AA-Video-I2V v2.1, just behind Dreamina Seedance 2.5
Utopai X is the video model from Utopai Studios, post-trained on MiniMax H3 and available inside PAI, Utopai's production platform, and now supports Image to Video!
Utopai X is the only model in the top 2 of both AA-Video-T2V v2.1 and AA-Video-I2V v2.1. On AA-Video-I2V v2.1, it is the highest ranked model built on MiniMax H3, ahead of MiniMax H3 (768p) itself, Hyperflow H3 and MiniMax H3 Max. In our testing, it does best on prompts that test Physics and Multi-Scene & Narrative, and on Frontier, UI/UX & Motion Design and Productivity & Knowledge Work use cases.
Utopai X is available in PAI at 93 PAI credits per second of 1080p video, with no public API. PAI plans start at $15 per month for 1,000 credits.
Congratulations to @UtopaiStudios on the release!
See below for our analysis of Utopai X in Image to Video 🧵
Mike Taylor@hammer_mt ·
Deep pull
- Artificial Analysis
@ArtificialAnlys ·
We've launched our HLE-Diamond leaderboard, our independent evaluation of the refined 1,000-question version of Humanity's Last Exam, as part of v5 of the Artificial Analysis Intelligence Index
HLE-Diamond was released by @CAIS and @scale_AI after a year of cleaning and refining the original Humanity's Last Exam dataset with input from the research community. It has 1,000 closed-ended questions, split evenly between reasoning and knowledge, and 272 of them include an image.
We run every question once without tools, using the official HLE system prompt. GPT-5.6 Luna (medium) grades each answer against the reference answer with the official HLE judge prompt, and we report the share of questions judged correct (pass@1).
Key takeaways:
➤ Only 17 model variants (~13%) score above 50% today: Gemini 4 Carbon (high) leads at 69%, followed by Claude Opus 5.5 (max) at 68%
➤ Models score higher on reasoning than on knowledge: Over 90% of the models we measured do better on the reasoning portion of the benchmark. Gemini 4 Carbon (high) leads both, at 80% on reasoning and 59% on knowledge. Claude Opus 5.5 (max) and GPT-6 Astra (max) both score 78% on reasoning, and Opus 5.5 is ahead on knowledge, 58% vs 50%
➤ Open weights models trail by more than 30 points: Kimi K3.1 Pro (weights not yet released) is the top open weights model at 37%, followed by MiniMax-M3.1-Flash at 28% and Kimi K3 (max) at 27%
- Gergely Orosz
@GergelyOrosz ·
A week from now, hosting two epic, free-to-attend events in the evening, in SF (registration required):
Wednesday (4 Nov): The Pragmatic Engineer Unplugged with @CompleteSkeptic & @bcantrill
RSVP: https://luma.com/pragmatic-engineer-unplugged
(cont'd)
nicksaraev@nicksaraev ·
AI has solved infinite personalized education.
This took $0.35 and ~15 minutes to make. You can create one on any topic, length, and level of complexity for a penny or two more.
Soon it'll be $0.00001 in tokens and realtime, next to you.
Human educational content is dead.
Max Stoiber@mxstbr ·
🫡
What should we ship next?
Ethan Mollick@emollick ·
Each of my posts has dozens of obvious and numbing AI slopreplies.
Worse, when humans really do reply (real example: someone from Anthropic addressing the issue I posted about), the "replies" algorithm does not surface them first
You can't really call this social media anymore.
Cognition@cognition ·
We've evaluated Gemini 4 Carbon and Gemini 4 Argon on FrontierCode 2.0, our coding benchmark that measure's a model's ability to go beyond the spec.
Gemini 4 Carbon scores near SOTA with a score of 40% and a cost of $4.04 per task, behind Claude Opus 5.5, GPT-6 Astra, and Claude Sonnet 5.5.
Gemini 4 Argon scores 26.3% with a cost of $6.41 per task, ahead of Muse Wonder 1.0, but behind Claude Haiku 5.5 and Grok 4.7.
See the leaderboard with full scores: https://cognition.com/frontiercode
Scott Wu@ScottWu46 ·
Really excited to work with this team!