AkashBytes @AkashNextGen

AI is moving beyond chatbots toward autonomous systems that can manage workflows, use tools, and execute tasks in the background. 🔥

Vaibhav Sisinty @VaibhavSisinty

2 days ago

The interview with @thsottiaux from OpenAI 🔥 0:00 Everything OpenAI launched at DevDay 2:44 Sponsors: Wispr Flow, Hostinger, OpenArt 4:11 What is ChatGPT Dots? 7:52 Will ChatGPT Dots be free? 10:15 GPT-6.1 Sol vs GPT-6 Astra 11:44 Is AI now building AI? 12:53 Ultra Fast:

69 60 540 233K 430

The biggest misconception about multimodal LLMs is that they somehow "understand" images. This couldn’t be further from the truth… A vision encoder converts image patches into embeddings aligned with text embeddings. After that, the LLM treats both as one sequence. Multimodality starts with aligned representations.

BREAKING: OpenAI just fired three of its own safety researchers for leaking confidential information to an outside AI safety org. This is not a routine HR story. OpenAI confirmed the terminations, saying the researchers "mishandled sensitive information outside established company procedures." The names, the org, and what was shared have not been disclosed. The timing is brutal: this comes two days after the NYT reported that OpenAI executives had repeatedly brushed aside internal safety warnings from employees. So the picture forming is: safety concerns raised internally, concerns dismissed, researchers go outside, researchers get fired. OpenAI is cutting ties with safety people while simultaneously being accused of ignoring safety people. That optics problem does not go away quietly. Is this a company enforcing policy, or silencing dissent?

Thanks @Forbes for featuring the great work we’re doing @ChronicleBioAI. There are exactly 0 FDA-approved therapies for the hundreds of millions of people suffering from diseases like POTS, Long COVID, MECFS or hEDS. We created ChronicleBio to change this. By building the deepest dataset on these conditions and leveraging AI, we will understand what the biology behind the symptoms is. From there, we can map the right drugs to the right patients.

Forbes @Forbes

2 days ago

Sufferers of complex, chronic diseases like POTS and long COVID-19 often spend years searching for relief. This former OpenAI exec started ChronicleBio to find treatments using AI and lots of biological data. Read more: forbes.com/sites/amyfeldm… 📸: Joel Saget/AFP via Getty Images

9 68 363 80K 132

If the New Tech Elite wants to compete with the Old Ruling Elite and take over The Center, they need to build in a new style that recovers civilizational things that have been lost, while invigorating a new vision of the future. This is what I call Greco-futurism.

keyboard_arrow_left Previous keyboard_arrow_right Next

Micah Springut @mspringut

2 days ago

If Florida actually wants to compete with the NE and CA for the best uni students they'd build tropical classical campuses.

keyboard_arrow_left Previous keyboard_arrow_right Next

35 41 1K 88K 170

I asked opus 5.5 for a motion design showreel and it labeled every shot like a portfolio 01 particles, 02 kinetic type, 03 wireframe, 04 flow field, 05 physics, 06 halftone, 07 data viz, 08 morph, 09 tunnel, 10 isometric 15 seconds. not one keyframe. it's all code

Florence @flxrnc

3 days ago

x.com/i/article/2105…

2 1 16 5K 13

$MU expects revenue to grow sequentially every quarter in FY27 and CFO says Q1 should be “the floor for gross margins” before margins improve through rest of year. Really great to see that even with price increases starting to cool that earnings can keep moving higher as mix and costs do more of the work.

Shay Boloor @StockSavvyShay

3 days ago

$MU says memory and storage will be “much tighter” in fiscal 2027 and 2028 than in 2026 which is an escalation from last quarter’s “tight beyond 2027.” Thats huge since a lot of the $40B+ capex is going into buildings and cleanrooms now while new supply needed to ease the

keyboard_arrow_left Previous keyboard_arrow_right Next

39 94 977 232K 81

Any AI model (in fact any tool) can be used for good or bad purposes. And this includes Anthropic's own models. Anthropic throwing shade at GLM-5.3 and other open weight LLMs can be interpreted in no other way than "You can only trust us with AI, others should be banned from it." Such narratives will only damage Anthropic's reputation. You need to regulate the use cases, not the model itself.

Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius

3 days ago

so Anthropic files for IPO and then immediately drops a blog post calling one of their biggest open-weight competitors a danger to the world? classic! now seems like a good time to point out that I built and run OBLITERATUS, an abliteration framework with over 8k stars and 1.5k

111 304 3K 119K 638

As Jaxson shows, "cheap LLM that grabs official docs" is trivial to set up. Like, literally any of us could do it in an afternoon. Hard work is 1) enter data only once a la Estonia, 2) clear simple forms for govt interaction. But yes, cleaning up ugly pages (CRA?!) is step one.

Jaxson Khan @jaxson

3 days ago

1/ The US launched America(.gov) yesterday and it got a bunch of buzz. It's an "AI front door" to 29,000 federal websites. Canadians can do the same or better! I was curious, so overnight I built Canada.ca 2.0 (with Claude) Ask in plain English or French,

7 12 95 15K 35

Your AI agent doesn't need a new browser. Give it the browser you already use. I found BrowserSkill by Tencent. It connects AI agents directly to Chrome or Edge and lets them work inside your existing browser sessions. That means your agent can: → Read pages → Search websites → Fill forms → Click and type → Upload/download files → Capture full-page screenshots → Work with logged-in websites → Debug websites and APIs And here's the wild part: You can keep watching it work. BrowserSkill gives the agent a separate visible Agent Window, while you can still take control whenever a human-only step appears. It can even investigate a website bug by connecting browser actions with: Requests + responses + console logs + page changes + performance data So instead of telling an AI: “Here are 10 screenshots. Figure out what's wrong.” You can essentially say: “Open the website. Reproduce the bug. Find the failing request. Inspect the response. Tell me what broke.” And it doesn't force you to use a specific AI model. It works with agents like: Claude Code Codex Cursor OpenClaw CodeBuddy WorkBuddy Pi Hermes Agent …and other shell-capable agents. The really interesting shift: AI agents are moving from generating instructions to actually operating the software we use. Your browser already has your accounts, sessions and context. BrowserSkill gives an agent a controlled way to work with it. The browser used to be where humans worked. Now we're building ways for AI agents to work there too. REPOO👇

1. OpenAI’s demo glitches got more attention. Meta’s new announcement got more attention. 2. OpenAI Dots requires OpenAI Pro, starting at $100/month. Muse connects to Shopify, QuickBooks and Stripe for small businesses and has a free tier. 3. OpenAI asking you to subscribe first. META is distributing Muse for free. 4. Dots gave the internet new memes. Muse gave users more reasons to stay. Unofficial scoreboard: $META 4 OpenAI -1

Anand @anandragn

5 days ago

OpenAI DevDay is tomorrow, Tuesday, September 29. Sam Altman’s keynote starts at 10 AM PT, streaming live. Their rumored lineup: 1. OpenAI o, an always-on agent, OpenAI’s answer to Meta Muse. 2. Astra 6.1 3. ChatGPT + Work, unified But the biggest announcement already happened

9 2 123 45K 15

$ORCL is up over 7% after OpenAI said enterprise is now its fastest growing business with B2B revenue up more than 100% since July. Thats huge because recurring enterprise inference makes it more likely OpenAI can grow into the massive infrastructure commitments already sitting inside Oracle’s backlog.

Shay Boloor @StockSavvyShay

4 days ago

OPENAI’S ANNUAL RECURRING REVENUE IS APPROACHING $70 BILLION

keyboard_arrow_left Previous keyboard_arrow_right Next

19 10 222 160K 11

images taken from @awodias's amazing article Current AI wouldn't be able to generate such detailed explanations, images, and flow. This person is definitely an expert or someone closely involved with the technology in question. (If sharing the image is an issue, please let me know and I will delete it immediately.)

keyboard_arrow_left Previous keyboard_arrow_right Next

awodias @awodias

3 days ago

See my article for a full breakdown on Ficontec🫡 awodias.substack.com/p/ficontec-bui…

1 1 8 9K 5

One diagram to understand how an AI agent works. The animation follows a single request: "Summarize Q3 churn, email the team." The agent goes around the plan, act, observe loop three times before it finishes. Lap 1: the LLM plans the work, calls SQL through the guardrails, and gets the churn numbers back. Lap 2: it calls Python to build the chart. Lap 3: it tries to send the email. The guardrails hold the call until a person approves it, then the report goes out. The dot colors show what is moving. Orange is the task and every command. Blue is results coming back from a tool. Pink is short term memory, read at every plan step. Yellow is long term memory, written at every observe step. Lime is a trace event heading to the observability panel. Six parts make this work: Instructions set the goal and the limits before the first step. The LLM chooses the next step and hands the action to a tool. Memory carries context between steps and between runs. Tools act on real systems: a database, a code runtime, an inbox. Guardrails check every tool call before it runs. Traces record the plan, each call, each result, and each approval. The approval rule I use is based on what happens if an action goes wrong. Reading data and drafting a chart can run on their own. Changing internal records gets a stricter check. Anything that leaves the organization, like an email to a team, waits for a person. Keeping that rule in the guardrail layer puts the policy in one place where reviewers can read it. Traces are what let you answer questions after the fact. If the agent sends the wrong report, the trace shows what it read, which step it picked, what each tool returned, and who approved the send. Where would you put the approval line for an agent that can read, write, and send?

OpenAI basically dropped an entire AI stack at DevDay 2026: 🟢 Dots:- always-on AI agents with their own cloud computer that can work in the background across connected apps. 🟡 GPT-6.1 Sol:- near-Astra-level performance for coding, computer use & professional work at a much lower cost. 🟠 Ultrafast:- a new high-speed inference tier, up to ~8× faster. 🔵 Pro 500:- $500/month plan with the highest usage limits + Ultrafast access. 🟣 Decisions API:- fast, constrained decision-making for classification, routing & choosing an agent’s next action. 🔴 Agents API:- build durable agents with hosted sandboxes, computer use, tool search, multi-agent workflows & context management. 🔵 Codex Cloud:- run coding tasks remotely with reusable environments that keep working even when your computer is off. 🟣 Codex Security Cloud:- scan codebases for vulnerabilities and help prepare fixes. 🟡 Plugin Extensions:- developers can now build interactive apps, panels and file viewers directly inside ChatGPT. 🟢 ChatGPT Spaces:- shared workspaces where people, AI agents and files can collaborate. 🟣 Pages:- collaborative, editable documents/workspaces inside ChatGPT. 🔵 Sign in with ChatGPT: developers can let users authenticate with ChatGPT and potentially bring their ChatGPT plan into apps. 🟠 @ChatGPT for Slack & Teams: bring ChatGPT directly into workplace conversations. 🟢 Private Intelligence: stronger privacy infrastructure for enterprise AI workloads. Which launch are you most excited about? Basically now ChatGPT is no longer just an AI chatbot. It’s becoming an entire AI ecosystem.

Nandkishor @devops_nk

4 days ago

Software Engineer's life after 10am PT.

33 128 3K 228K 246

Update to the Agent Arena Pareto frontier: GPT-6 Luna (Max) has landed on the Agent Arena Pareto frontier with +1.59% net improvement at a median cost of just $0.05 per task! Luna costs 29% less than DeepSeek-V4.1-Flash and 76% less than Hy4 preview, while trailing their net-improvement scores by just 2.27 and 2.51 percentage points respectively. Compared to other GPT models on the Pareto frontier, GPT-6 Luna (Max) costs 94% less than GPT-6 Sol (Max), and 98% less than GPT-6 Astra (Max). Congrats to the @OpenAI team on this cost-efficient release!

Arena.ai @arena

5 days ago

GPT-6 Luna (Max) by @OpenAI is #23 in Agent Arena with +1.6% net improvement across 8K real-world agentic sessions from our global community of users. Although GPT-6 Luna (Max) did not land on the Agent Arena Pareto frontier, it remains a cost-efficient model. Its $0.05 median

20 21 393 107K 45

🚨 Sonnet 5.5 is looking REALLY scary Up to 30% lower cost per task than its predecessor 30%+ faster output generationFastest Sonnet model yet And the crazy part: Same pricing as GPT-6 Sol, but significantly faster and more efficient while looking MUCH stronger overall from the outputs I've seen Anthropic is seriously back

Salio @Mr_Salio

5 days ago

🚨 Sonnet 5.5 Benchmark dropped It smashed Opus 5.5 and GPT-6 Astra Sonnet 5.5 mogging best openAI model wasn't in my checklist 😂

13 19 341 45K 35

What life with AI will look like in one year:

BuBBliK @k1rallik

5 days ago

This shouldn't be FREE - AI that turns ONE image into a world you can explore Move through a doorway, look around, then turn back. The WILD part? It helps remember what was there. WorldCrafter, from Tencent ARC Lab and Peking University, generates video as you move the camera.

5 2 31 10K 20

NVIDIA's CEO has a blunt message for the AI industry. "If the product is not safe, just don't release it. It's that simple." Jensen Huang says if Anthropic and OpenAI keep warning the world about the dangers of their models, they should stop shipping them until those risks are addressed. But he also questions whether they truly believe their own warnings: "I really don't believe they mean that. I really don't believe that they're building something they have no idea how to control." His approach isn't to slow down AI development. It's to build better containment and monitoring systems, faster. The CEO of the company supplying computing power to AI labs is openly questioning their safety warnings. What do you think about all this?

You cannot prompt an ASR system out of missing acoustic coverage. Eventually it has to hear the language, accents and conditions it needs to handle.

Shobhit Banga @shobhitbanga

5 days ago

Introducing Monsoon ASR ⚡️ @voicearena_ai Speech recognition does not have a model problem anymore. It has a data problem. The best ASR systems are approaching human-level performance in English. But move into the long tail of the world's languages, especially real,

21 78 279 638K 172