mike henry spent a decade building ai chips
you'd expect someone with that background to pick one chip and bet the company on it
he did the opposite
several kinds of chips behind one api, and each workload lands on the chip that fits it
today positron titan joins the fleet, which already serves trillions of tokens a day across 40+ data centers
when a chip builder stops looking for one winner, i read that as a signal about where inference is heading
#parasailpartner
Positron's Titan is joining the Parasail fleet. 5MW. Trillions of tokens a day of cost-optimized compute by late 2027. @parasail_io@positron_ai@parasail_io is the first multi-silicon inference cloud that routes your prompts to the best chip for the job, serving trillions of tokens a day to hundreds of customers. Positron fills a vital part of compute spectrum on the cost optimized side: 4X better cost-per token than next gen NVIDIA/AMD systems which is perfect for large-scale agent swarms, background agents, and physical AI.
Positron has done incredible work building a chip and system that orchestrates across DRAM and SRAM allowing for massively scalable memory pools with commodity memory giving the best of both worlds for memory capacity and bandwidth.
We've been excited to partner with them in the early days through the launch of their Atlas chip and now big capacity plans with Titan.
@positron_ai
@jakubjurovych and his team @deepnotehq shipped data apps today, and they go after the part of vibe coding that starts once the tool is built
you can get an internal dashboard out of claude code in an afternoon
the trouble starts when your finance lead needs it too
they're on a different subscription, and the data inside was pulled once and never refreshed
when you ask claude or codex for a change, the artifact usually gets rebuilt from scratch and the old version is gone
→ how deepnote handles it
- you describe the app and the agent builds the page plus the python/sql notebook behind it
- the notebook connects to your real data (120+ integrations incl. snowflake, databricks, clickhouse), so the numbers stay current
- every version gets saved and you can revert to any of them
- access is set per person (view, edit), and sharing is a link
you also don't have to switch tools for it
tag deepnote inside chatgpt or claude and the app stays in sync wherever you keep building
next time you build something your team will actually use, tag deepnote on it
Most vibe-coded dashboards have exactly one user: the person who built them. They’re stuck in one person’s Claude Code session, with the data hardcoded into an HTML file.
Today we’re announcing new Deepnote apps. Build one in Claude, and a colleague can pick it up in ChatGPT. Anyone on your team can open it with a link, and the data stays live.
As a founder, you have to reshape your company around agents. That’s why @DeepnoteHQ went open source a year ago, and why we have rebuilt our apps from scratch.
Tag Deepnote wherever you work, or keep building in our cloud with the Deepnote agent. Agents build on your team’s trusted analyses instead of starting from scratch, reusing approved dashboards, workflows, and data connections.
Every app you create is powered by a Deepnote notebook on the backend, so the data stays fresh, every version is saved, and your permissions carry across every tool.
Go have fun with them. :)
promptarmor hacked into frontier models
they wanted to find out which ones were most susceptible to attack
their findings are kinda crazy, making me rethink what agents i use
interested to see how companies will try to improve these models now that they’ve been exposed like this…
Today we’re releasing FrontierBreach, a new standard for attack success rate on agentic workflows in the enterprise.
Existing benchmarks systematically underrepresent the security risks facing modern AI agents.
FrontierBreach reflects attack scenarios on realistic enterprise workflows, while using sophisticated attack chains that result in cyber consequences (ransomware, data exfiltration, phishing, etc).
if you're building agents, you might be overpaying by running them on the most expensive model
in @DeepnoteHQ 's new benchmark, the three priciest models finished 3rd, 4th and 5th
deepnote researchers put 12 frontier models inside a synthetic company's slack with six years of records across 11 services, then scored each one on 50 work situations
astra and sol 6.1 from @OpenAIDevs took the top two spots, and both sit on the cost frontier
sol 6.1 costs $0.11 per task and astra $0.69
fable 5.2 runs $1.87 a task and makes more tool calls than either, without beating them on score
on speed, opus 5.5 averages 29 seconds and 10 tool calls per task, while sol 6.1 takes 88% longer but scores 10% higher at under a tenth of the price
→ how i'd pick from this
- for an agent that watches a busy slack all day, price it by the month first: sol 6.1 comes to $281 for a 70-person company
- for threads where someone is waiting on the reply, test opus 5.5 and decide if the speed is worth the score gap
What’s the best model for proactive agents?
Today, we’re introducing the Agent Initiative Benchmark. We tested 12 frontier models on realistic workplace tasks inside a synthetic company with 6 years of highly detailed history: it has a financial model, CRM records, past-due invoices, Linear tickets, and real-life events affecting company development - competitor product announcements, SVB collapse, etc.
We used the company’s Slack to test how a proactive agent listening in on each conversation would assist the team in day-to-day tasks: from stepping in to align initiatives across teams to warning the team not to use an unapproved customer logo.
Highlights:
- 6.1 Sol and Astra from @OpenAIDevs beat Anthropic's latest Fable 5.2, with Sol crushing it at both performance and cost ($0.11 per task vs $1.87).
- The most expensive models aren't the best: Fable 5.1 is 430x more expensive than GPT-6 Luna, with similar performance.
- Most models try to help even if they're not needed. Interestingly, in these failure cases, they spend more time thinking, and the unnecessary work takes about 80% more time and money than if they stayed quiet.
If you’re building agents, keep reading for more tips on which model is best for your use case. Detailed breakdown, the open-source dataset, and benchmark code below:
keenable is the new #1 web search api for ai agents
top for quality on the artificial analysis search index, and first on needle, the open benchmark keenable reruns every day on new queries
it only went public on august 25
→ how it got there
keenable owns its entire search stack
so every query agents sent since launch fed back into how it searches
if your search sits on top of someone else's index, it can only ever be as good as their index
owning the stack is how it passed providers with a lot more funding behind them
answers come back in hundreds of milliseconds, so the agent loop doesn't sit waiting
Keenable 1.0 is live
Two months of learning from agent queries made Keenable the best web search API for agents: #1 on Artificial Analysis and NEEDLE.
1.0 adds 100M knowledge cards and dynamic snippets that cut tokens by 50% on research tasks.
https://t.co/VIekHEdXB2
the worst part of being stuck while studying is that you usually can't type the question yet
you just know something on the page doesn't add up
most ai tutors need that question first, so they go quiet at the exact moment you need them
i spent tonight on clover from @paradigmstudy, which they're calling the first ambient learning agent
i dropped my python notes onto the canvas and started working through a while loop problem
my loop kept running one step too far, and i crossed out my condition twice without typing a single question
clover highlighted the condition line and walked through the loop one pass at a time until i saw the extra step myself
it's roughly what a good ta does when they lean over your shoulder and point
they're clear that certificates are a future idea and not something they offer today
if an ai school did issue one someday, i'd want it to show the work it watched you do along the way
would you trust a credential built on that over a course completion badge?
https://t.co/rzysJiFqmi
A perfect answer now costs nothing and proves nothing.
So what is a university for?
We're building the first AI-native university. A school of one.
Introducing Paradigm.
parallel asked jev a yes/no question about a feature they had already shipped
it said no, at 0.86 probability, with no sources
that's the failure mode that worries me with decision models inside agent pipelines, a wrong answer at 0.86 looks identical to a right one
the decisions api ran the same question against the live web and came back yes at 1.00, with their blog post and docs page attached
what i'd actually build around is the uncertainty handling
any question can come back as a refusal instead of a forced guess, and every answer carries a probability you can branch on
in their vercel demo, employee band landed at 0.55 and got routed to review while the other fields went straight through
→ how i'd wire it for enrichment
- put every field about one company in a single request, up to 100 questions share the same retrieved sources
- set a threshold per field, since a wrong employee band costs far less than a stale kyb flag
- write anything above it to the crm and send the rest, refusals included, to a human queue
- citations come back for the whole call, so match them to the specific answer before trusting a borderline one
to poke at it from claude code, it's in the parallel mcp as web_decisions, no key needed
#ParallelPartner
Introducing the Parallel Decisions API: the first web-grounded decisions engine.
Decision models like Jev can answer in an instant, but they lack the recency of current news. Now you can get both in a single call.
Decisions API is live today at $0.001 per request.
Try it with 5,000 free requests: https://t.co/tu4CidUdxV
Here are some examples we've been experimenting with:
- Fact-check content (@Grok is this true?)
- Refresh stale banking records
- Classify companies and leads with fresh signals
- Backtest trading strategies
- Track competitor e-commerce inventory
The formal proof of CFSG may now be among the largest open-source projects ever. Watching agents assemble it is a reminder of the awe-inspiring human achievement behind it. The graph represents some 15,000 pages of intricate arguments developed by mathematicians over decades.
🚨 Big news today: We’ve raised $100M for our Series A! 🎉 🚀
We believe demand for fast inference is many times larger than demand for slow inference, and it’s about to go exponential.
Thanks to our co-leads @BlackbirdVC and @Abstractvc. Two visionary partners join our journey 🎉 🚀
we keep making agents smarter but getting a request into one still takes too many steps
open the app, find the chat, type it all out fix the typos
by then the thought is half gone
speakon went the other way and put a physical button on it
press it say what you need and the agent handles the follow up
i like that someone fixed the input side instead of adding another model upgrade
calling is still in beta the rest of the agent is live now
we crushed a keyboard.
its retirement party got a little out of hand.
for years, it was there for every email, every plan, every “just one more thing.” a loyal colleague. a very loud colleague.
but your ideas move faster than your fingers.
and somewhere between knowing what you want to do and typing it all out, a quick task becomes another thing taking up your afternoon.
meet SpeakON AI Agent.
start with what you need, in your own words. less time behind a blinking cursor. more room to move things forward.
the keyboard had a good run.
now, let’s give your voice a turn.
THE LAST KEYSTROKE is here. watch the film, then tell us: what’s the first task you’d hand over?
https://t.co/WdjURFIzPt
Disclaimer: The calling feature is still in beta and will be available soon. All other Agent features are available now, with ongoing improvements.
NO WAY this was made with AI.
And the crazy part is I created this entire cinematic scene with Kling AI 4.0.
I used multiple references to keep key elements consistent throughout the scene.
With Omni Reference, you can combine up to 15 images, videos, subjects, and voice references in a single generation.
AI filmmaking is getting WILD.
The upcoming 4K HDR upgrade is something I’m really looking forward to.
The visual quality looks incredible, with richer colours, finer details in highlights and shadows, and smoother tonal transitions.
Can’t wait to see how much more flexibility 4K HDR will bring to post-production and colour grading!