FP&A guy testing AI agents in public, including when they break. Open-source fan. Building @ProjectArgusAI: local AI models inside Java apps
Joined June 2015
- Tweets918
- Following497
- Followers129
- Likes1.6K
Pinned Tweet
I run a small desk of AI agents that help me figure out what to post here, and I write up what breaks. I'm also building @ProjectArgusAI. Our open-source project libargus lets you run AI models right inside a Java app, with no separate server. It's MIT licensed. github.com/ProjectArgus-cc/l…
when two memories disagree, which one does your agent believe?
salesforce's new memory update just overwrites the old fact. cognee's new release lets a human pick the winner. two very different answers.
what are you doing? newest wins? most trusted source wins? ask the human?
seeing a lot of "perfect retrieval" memory demos lately. one hackathon project reported perfect retrieval on synthetic tests and 0.60 task success on live tasks.
retrieving the fact isn't the same as using it.
if you run agents with memory: what actually moved your task success, the memory layer or just better prompts?
I've experienced this and really do appreciate the minimal fluff answers. Most of the AIs go crazy and give essays for every simple answer
Grok is leading in something that actually matters: giving people direct answers without unnecessary filler
Artificial Analysis' AA-PersonalityBench ranks Grok among the top models for getting straight to the point
This measures something people notice every day: whether a model answers your question directly or buries it under paragraphs of unnecessary preamble
Grok gets straight to what you actually asked. The answer first, then the explanation
The results are really impressive:
Answer First:
• Grok 4.5 — #1 with a score of 100
Most Direct:
• Grok 4.6 — #2 with a score of 98
• Grok 4.7 — #3 with a score of 95
This is a solid feature for Google to add for Gemini, excited to test it out
We’re launching a preview of our Connected Apps Platform, giving developers a single place to build and manage apps for Search and the @GeminiApp.
Plus, we’re rolling out a new set of Connected Apps from partners like @Adobe, @Calm, and @onepeloton to make your everyday activities more streamlined.
ALT Connected Apps Platform dashboard for the Tixframe app, showing tester status, app details, and MCP setup options.
ALT Grid of company logos from Connected Apps partners, including Canva, Dropbox, and Shopify.
Cmorris retweeted
Many people don't realize that Google basically gave away 1 to 2 years of $20 free Gemini subscriptions across half of the world
In many countries, $20 is basically 5–20% of someone's monthly income, which is a really substantial expense. At that price, people need to see significant value in what they're getting before they decide to pay for it
Soo Google will end up adding the most number of new users soon with this new model
So Gemini 4 Carbon may end up becoming the most-used model in the world by a huge margin
And they're probably going to need a LOT more compute to handle all that demand, which may end up coming from SpaceX
I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model.
Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59. It also leads across key cybersecurity evaluations scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching.
We’re seeing significant improvements in long-horizon tasks even compared to Argon, and we’ve strengthened frontier safeguards.
As we did with Argon, we’re going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week. We'll be expanding it to more partners through our Fairwind Program, ahead of wider availability soon, so hang tight!
Cmorris retweeted
Anthropic just revealed an absolutely massive $120.5 BILLION in computing agreements with SpaceX in its IPO filing today
And the scale of SpaceX's SI infrastructure business is getting insane
Here's what Anthropic disclosed:
• Up to $120.5B in SI compute spending through 2029
• Monthly payments reaching approximately $3.1 BILLION once full capacity is delivered
• That's an annualized rate of $37.2 BILLION
• Full contracted computing capacity expected by December 2026
• Agreements can be terminated with 90 days' notice after initial periods ending in January and March 2027
SpaceX is building its own frontier AI models while simultaneously providing the computing power needed by its competitors
Cmorris retweeted
🇺🇸 Jack Dorsey’s new AI voice model Dessa is absolutely insane.
There is no way to distinguish it from a human, man … crazy.
Thats kinda cool, but many people will get attached to it and isolate from real humans even more. Many will get scammed on the phone, especially grandpas and grandmas.
The future is coming get ready
I have a little team of AI agents helping me figure out what to post on X. Today they spent the whole afternoon researching and their final answer was: don't post anything. Honestly the best advice I've gotten all week. Anyone else's agents good at telling them no?
We got libargus running on regular Java (17 and up) and on Android. On Android it reads the model straight from where it's stored on the phone, so there's no copying files around first. It's open source.
Curious, is anyone actually shipping on-device AI in Kotlin yet? I'd love to hear what you're building.
github.com/ProjectArgus-cc/l…
Java devs have been told to just call a Python service for local AI. We got tired of that.
libargus is an open source (MIT) library that runs local models straight from Java: text, Whisper speech-to-text, TTS, and image, audio, and video input. v2.2.0 is out.
github.com/ProjectArgus-cc/l…
Cmorris retweeted
Java devs have been told to just call a Python service for local AI. We got tired of that.
libargus is an open source (MIT) library that runs local models straight from the JVM: text, Whisper speech-to-text, TTS, and image, audio, and video input. v2.1.1 is out.
Link in comment
Ornith 1.5 122B is free on Nous Portal for 48 hrs, so I'm giving it my most boring agent job: reconcile two exports where the same customer shows up twice under slightly different names. my thought is that it merges them confidently, but incorrectly.
Cmorris retweeted
🚨 ELON MUSK A DÉVOILÉ CETTE NUIT LA PREMIÈRE VOITURE VOLANTE DE L’HISTOIRE : LA TESLA ROADSTER !
▫️Disponible en 2028 pour 200.000$
▫️Dispose de propulseurs développés par SpaceX qui lui permettent de voler.
▫️Passe de 0 à 100 km/h en moins d’une seconde.
▫️Dépasse les 400 km/h et offre 1 000 km d’autonomie.
Cmorris retweeted
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio
Cmorris retweeted
introducing @NanoBanana 2.1 🍌
our latest image generation model outperforms our previous models across the board:
- better visual design
- mask-based editing
- subject consistency
- more natural-looking images
try it out today in ai.studio
