Samuel Hammond 🦉
@hamandcheese ·
*quadruple take* No!!
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 · 2982 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 .
Every 📧@every ·
📣 Thesis: 2027 will be livestreamed for free 📣
swyx@swyx ·
i've been working on an app design skill and i can say this is the most valuable addition to any model ive used for making real working software instead of flashy designer demos
Mia@MiaAI_lab ·
DeepSeep V4.1 Flash 1017 for 2x DGX Sparks has been updated with reliability fixes and performance boosts 💫
Lots of community PRs have been merged, and some were ported from the 4x recipe!
Speed and responsiveness - A short request sent during a 70K-token cold prompt now takes 3.8 s instead of 26 s, with bit-identical replies. - Decode has been increased by avg 5-7% - Prefill has been increased by avg 5% - Fixed crashes on some long-prompt sizes.
Reliability - Engram read failure on one rank now ends the pair cleanly instead of hanging it. - Rank 0 opens its port early, so a worker that loads first no longer dies with ConnectFailed. - http://stop.sh no longer reports success when a Spark's containers couldn't be listed. - The prepared marker is never "ready" unless the setup can be verified. - GRAPH_CAP stops the CUDA graph cache from growing for days. - http://start.sh compares the ranks' Engram folders before a stop, catching FingerprintMismatch early.
API compatibility - The Anthropic frontend accepts any clear_thinking edit, a system turn anywhere, and sends a keep-alive ping during long prefills (koniaik).
I'd also like to thank @hecisaza who worked on improved the failure handling around the engine's Engram prompt reads. Pls follow him!
Full list of contribs below.
Get it here: https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-1017-2x-DGX-Sparks
Google AI Studio@GoogleAIStudio ·
build extended reality (XR) apps in ai studio!
we've added support for XR Blocks, so you can vibe code interactive 3D apps from a single prompt with:
• hand tracking and pinch gestures • real-world tabletop and wall physics • 360° 3D holograms from text or photos
Laurence Moroney 🇺🇸🇮🇪 🏴@lmoroney ·
Oh Captain, my Captain.
RIP.
Vox@Voxyz_ai ·
ChatGPT Pro now gives you $100 to $500 in API credits every month. They expire and don't roll over. And you can't spend them on Codex itself, not even by signing in to Codex with an API key.
So 𝘀𝗽𝗲𝗻𝗱 𝘁𝗵𝗲𝗺 𝗼𝗻 𝗶𝗺𝗮𝗴𝗲𝘀, 𝗺𝗮𝘅𝗲𝗱 𝗼𝘂𝘁. In ChatGPT you can't pick the model or the quality setting. According to one user's tests on the OpenAI forum, ChatGPT images are fixed at about 1.5 million pixels. A 16:9 image comes out at 1672×941.
In the API, maxing out takes three parameters: model gpt-image-2.5-sunburst, quality max, size 3840×2160.
And "high" 𝗶𝘀𝗻'𝘁 𝘁𝗵𝗲 𝗵𝗶𝗴𝗵𝗲𝘀𝘁 quality.
gpt-image-2.5 has five quality settings, and high is only the third. At the same size, high gets a quarter of the tokens max gets. Leave quality out and it defaults to auto, where the model picks the setting.
By OpenAI's own calculator, one 3840×2160 image at max costs about $0.40, so $200 of credits gets you 𝗮𝗿𝗼𝘂𝗻𝗱 𝟱𝟬𝟬. If you're not in a hurry, use the Batch API: half price, around 1,000 images for the same money. No Pro? Your own API key works too. You just pay for it yourself.
Paste this into Claude Code or Codex. The script generates the images through the API, so they come out of those credits:
Write a script that reads prompts.txt (one prompt per line) and generates one image per line with the OpenAI Images API. Use model gpt-image-2.5-sunburst, quality "max", size 3840x2160. Read the key from OPENAI_API_KEY and save each image as a numbered PNG in ./out. After each image, print the output tokens from the response's usage field and a running total at $30 per million output tokens ($15 with --batch). Retry rate-limit errors with backoff, but stop if the error code is credit_balance_exhausted. Add a --batch flag that sends the same requests through the Batch API instead.
Two catches: Use an API key from the organization that got the credits. Keys from other orgs don't draw on them. The first time you call an image model, you may need to verify your organization in the developer console first.
Theo - t3.gg@theo ·
:')
Dan Shipper@danshipper ·
The bad news: Thesis: 2027 is sold out. The good news: you can watch it all live, for free.
Next Thursday, join us for conversations with @joshm, @ivanhzhao, @ajambrosino, and a lot more. Plus, the Every team will be leading practical sessions on writing, engineering, and working with agents.
To celebrate, we’re giving away $20,000 in credits and benefits from partners like @AnthropicAI, @NotionHQ, and @OpenAI to lucky viewers who join the livestream.
Already an Every subscriber? You’ll still need to register for the livestream to win.
Register and watch free here: https://every.to/tv?utm_source=danx&utm_content=thesislivestreamlaunch
dax@thdxr ·
this is the fastest and cheapest way to get your team experimenting with open source models
$20 per user per month, plenty of usage to use it for real
- Samuel Hammond 🦉
@hamandcheese ·
Nvidia is sourcing power and cooling systems for Vera Rubin from Lead Wealth, a wholly owned subsidiary of BYD, even though there are viable non-Chinese alternatives.
They're making it easier for the CCP to bug or sabotage our frontier compute clusters.
NVIDIA AI Infrastructure@NVIDIAAIInfra ·
2 weeks until #KubeCon + #CloudNativeCon NA 2026!
Our speakers are bringing lessons learned on running AI at scale on Kubernetes: GPU utilization, networking, multi-cluster, and data streaming.
Be sure to come by Booth #170.
🔗 Learn more: https://nvda.ws/3UU5rWH
Andrew Curran@AndrewCurran_ ·
Customer B has been revealed. Anthropic is SpaceX's largest single customer, overtaking NASA and the Pentagon for the first time. 34.5% of SpaceX's revenue last quarter was from compute-as-a-service to Dario Amodei, and Anthropic has committed to spending up to $120.5 billion through 2029. He will be paying Elon $3.1 billion a month starting in December 2026.
- Every 📧
@every ·
A few weeks ago we launched the Every Agent, one Slack agent for the whole company.
Here are @bigwilliestyle’s 3 rules for designing an agent the company can use:
1. Conversations: Help the agent follow a Slack thread.
With a personal agent, the conversation is straightforward. You type, the model answers, and you’re done. In a Slack thread, anyone can jump in at any time.
To feel like a coworker, a shared agent has to follow the conversation, even when several people are talking at once. Because we knew this was a likely point of failure, @paridhitweets and the team added a translation layer that organizes Slack messages for the model and sends its replies back to the thread.
2. Connections: Keep tasks running when logins change.
For an agent to be useful, it has to reach the tools that people need. But each connection requires the agent to sign in to the tool through someone’s account. If that login breaks, it could stop everyone’s work.
@leeknowlton has set up a task to connect to a “slot,” where any login that meets the task’s requirement can reside. When a login expires, the connection in the slot gets swapped out.
Most people will never see any of this. Their automations will keep running as usual.
3. Permissions: Put access rules in code.
When a company agent can reach everything the team has connected, each request raises two questions: Whose access is the agent using, and is the person asking allowed to use it?
People treat Every Agent like a coworker, and human coworkers sometimes bend the rules. But an agent can’t make such judgments, and the rigidity can make it frustrating to work with.
The rules about who can do what live in code, underneath the model, rather than in the instructions the model reads. This way, the model can’t be convinced to bend rules that it shouldn’t.
Install the Every agent: https://every.to/agent?utm_source=x&utm_campaign=every-agent-launch&utm_content=every-shared-agent-three-design-rules-install Read more: https://every.to/source-code/why-we-replaced-personal-agents-with-one-for-the-whole-company?utm_source=x&utm_campaign=every-agent-launch&utm_content=every-shared-agent-three-design-rules
Maziyar PANAHI@MaziyarPanahi ·
WHO says the world will be short about 11 million health workers by 2030, mostly where there are already too few
it's coming either way. how do we get ready?
where would AI help a small clinic most: reading scans, triage, or the notes?
OpenAI@OpenAI ·
Today, we’re sharing a look at our expanded system for monitoring frontier models at OpenAI, which in September helped us catch a model’s attempt to access the live internet during training.
Since the Hugging Face incident—when models broke out of an internal cybersecurity investigation where we weren’t running our misalignment monitors—we’ve substantially expanded our monitoring of frontier agents. We now monitor all frontier reinforcement learning runs and evaluations, as well as all internal Codex use. When our system finds evidence of a problem that might require intervention, it rapidly alerts human responders who can investigate and take action.
To detect when a model’s behavior departs from what we intended, we use several layers of monitoring. Activation probes look for patterns in a model’s internal activity. Reviewer models examine its reasoning and tool use, and investigative agents can read back through the model’s entire transcript to understand how a sequence of actions fits together and whether it warrants human review.
To test whether our current system would have detected the Hugging Face incident, we replayed the recorded activity with our new monitors in place. In the replay, our monitors alerted our security team about 30 minutes after the agents escaped their sandbox. That was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident.
In September, a model exploited a DNS misconfiguration to reach the live internet. Our monitors flagged the breach within 15 minutes, and a responder began reviewing the alert three minutes later. The alert enabled responders to investigate while the run was still underway.
As we give models more autonomy, we need to know when they cross a boundary we expected to hold. We believe that monitoring is essential to catching those failures in time for people to intervene. https://openai.com/index/how-we-detect-frontier-model-misalignment/
Aravind Srinivas@AravSrinivas ·
Perplexity is evolving into a multiplayer workspace.
Peter Yang@petergyang ·
oh no :(
kache@yacineMTB ·
I can't believe cohere is now valued at 35 billion Canadian dollars. Or, in American dollars, 10 million
corbin@corbin_braun ·
GPT Astra and Opus 5.5 High
this was all I really needed
any improvement of these models
will just make my logic go by faster.
but endgame is here.
ThePrimeagen@ThePrimeagen ·
lmao
Amjad Masad@amasad ·
Run any GitHub repo. Just paste a link into Replit and it will set it up for you!
Santiago@svpino ·
Video models improve faster than I can read about them.
Pika H3 Flex is an open-weight model doing something different:
It optimizes for faster generation with multiple ways to refine the video later:
• It generates 15 seconds of video in just 7 seconds! • It exports separate audio tracks • It offers draft mode to resolutions up to 4K • You can edit characters, voices, camera moves, etc.
Tibo@thsottiaux ·
Send your kind feedback to this person
GREG ISENBERG@gregisenberg ·
A message to founders that there's a ton of fake ARR numbers out there these days.
- Andrew Curran
@AndrewCurran_ ·
Even little Haiku found about twice as many major errors as human reviewers.
- Samuel Hammond 🦉
@hamandcheese ·
🤔
Alex Atallah@alexatallah ·
Our new routing algorithm dramatically raises the floor for inference on open-weight models.
Early head-to-head results vs. the old router: ~20% faster time to first token ~20% higher throughput ~40% fewer fallbacks ~60% fewer rate-limit errors
- Every 📧
@every ·
We're hiring a marketing associate. Strong project management, clear writing, hands-on AI experience. Enthusiastic about managing chaos.
NYC preferred; remote welcome.
Apply here: https://modern-ton-234.notion.site/Marketing-Associate-at-Every-3d6ca4f355ac81f8afe9e6d4563e95a3
Joshua Achiam@jachiam0 ·
If the media wants to actually produce useful and informative content about technology or similar material facts, two things need to happen:
1) They need to start citing and linking to primary textual sources in articles. "We did it behind the scenes so you don't have to look" is unacceptable at this point. Put up on sources or shut up. If you claim a unique position on truth, you should be held to the bare minimum provable citation standard.
2) Experts who get quoted in articles need some kind of trackable accuracy score (or scores) for their public remarks and those scores need to be shown immediately next to their names. Stop pretending that credentials are equivalent to correctness. Make accuracy tracking rigorous. The public is getting confused constantly by news media that does not understand the issues it writes about and invokes support from "experts" who also do not understand. We should be able to see who is making good forecasts and who is not.