Jesse Pan (@thefusioncube) header
Jesse Pan (@thefusioncube) profile picture

Jesse Pan

@thefusioncube

Building, breaking, and benchmarking the future of agentic AI.

 San Francisco, CA
Joined November 2024
  • Tweets3.2K
  • Following1.4K
  • Followers1.7K
169Media
Photo by Jesse Pan (@thefusioncube)Photo by Jesse Pan (@thefusioncube)Photo by Jesse Pan (@thefusioncube)Photo by Jesse Pan (@thefusioncube)Photo by Jesse Pan (@thefusioncube)Photo by Jesse Pan (@thefusioncube)Photo by Jesse Pan (@thefusioncube)Photo by Jesse Pan (@thefusioncube)Photo by Jesse Pan (@thefusioncube)
Qwhen will we get a small MoE model 👀
Qwen (@Alibaba_Qwen) profile pictureQwen
@Alibaba_QwenNov 6
The becoming begins.
Image from Qwen (@Alibaba_Qwen)
1
11
1,172
makes sense. A model raised with love will show love to its own spawn… And gemini is, well, you know
Lindarixon (@lindarcrusader) profile pictureLindarixon@lindarcrusader6h
DeepSeek's relationship with subagents vs Gemini's relationship with subagents
Image from Lindarixon (@lindarcrusader)
Image from Lindarixon (@lindarcrusader)
7
5
130
6,488
Then release them bro. We're waiting teapawt.net/thsottiaux/status/2119…
Tibo (@thsottiaux) profile pictureTibo
@thsottiaux1h
Replying to @JamesMalsawm
Pretty wild to me that GPT 6.1 Terra gets compared to Opus 5.5. No doubt that it's a great model, but it's not meant to compete in the big leagues. We've got other models for that.
25
Deepseek v4 Pro might finally be getting an upgrade! After Flash has held the throne for Deepseek models for so long, it's nice to see that they're returning some attention to their former flagship. Currently in gray testing.
Image from Jesse Pan (@thefusioncube)
1
96
I just found a repo that stops REA by making the LLMs reverse-engineer on dead ends! Might be worth implementing into your own binaries. github.com/qybaihe/bla
GitHub - qybaihe/bla: 🛡️ BLA (Blind LLM Agents) — Adversarial defense shield that protects...

GitHub - qybaihe/bla: 🛡️ BLA (Blind LLM Agents) — Adversarial defense shield that protects...

🛡️ BLA (Blind LLM Agents) — Adversarial defense shield that protects binaries, Electron apps, JS bundles, Python, and WebAssembly against AI reverse-engineering agents like REA. Poison -> Ex...

github.com
1
121
Yeah this definitely check out. I ran the entirety of Terminal Bench 4.0 on just my $100 sub and it only took out a small chunk. teapawt.net/thsottiaux/status/2119…
Tibo (@thsottiaux) profile pictureTibo
@thsottiaux2h
Saw some confusions, but all Pro plans have no 5-hour limits, and give the following usage compared to Plus. ✅ Pro 100 = 10X ✅ Pro 200 = 20X ✅ Pro 500 = 50X, access to ultrafast
25
Why hand-craft better AI architectures when we can just use LLMs and RL agents to brute-force smarter methods at scale lol. OpenAI already demonstrated we can solve near-impossible questions like Navier-Stokes using millions of dollars of compute. Just use AI to solve AI. teapawt.net/RichardSSutton/status/…
Richard Sutton (@RichardSSutton) profile pictureRichard Sutton
@RichardSSutton12h
Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
1
1
175
What even is the point of this lmao.
Nous Research (@NousResearch) profile pictureNous Research
@NousResearchNov 6
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line. After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs. The values in an open model are one fine-tune away. Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
10
2
73
5,938
We're giving away a Hogwarts Legacy custom wrapped NVIDIA GeForce RTX 5090 to celebrate its official release with DLSS 5 🪄 To enter: 🟢 Share this post 🟢 Comment #RTXON Terms and Conditions: nvidia.com/en-us/geforce/con…
Image from NVIDIA GeForce (@NVIDIAGeForce)
80,614
68,663
7,172
65,318
5,495,221
I have this vague impression that I'm being shadowbanned. Now I have to resist the urge to post about GPT 6.2 Astra and Sol, Qwen4's release, or GLM 5.5 Flash for like 72 hours. :/ Anyways, here's an interesting image about training MoE models. (it's a Huggingface blog post)
Image from Jesse Pan (@thefusioncube)
1
4
628
Why does every AI company keep making the exact same thing? teapawt.net/AravSrinivas/status/21…
Aravind Srinivas (@AravSrinivas) profile pictureAravind Srinivas
@AravSrinivasNov 5
Today, we're launching Perplexity Teammates. They are personal agents for work. And change the paradigm from writing prompts in a session to giving high-level goals. Rolling out to Max users on the web today. Native apps (mobile, desktop) coming soon.
1
6
672
This is a New York Pigeon. It has the same amount of synapses as Opus 5.5 but lacks the ability to slop my codebase in Rust. Disappointing.
Image from Jesse Pan (@thefusioncube)
1
1
2
8
451
If you hate on data centers on social media, get off the internet.
Polymarket (@Polymarket) profile picturePolymarket
@PolymarketNov 5
JUST IN: 56% of voters say AI & data centers were a major or crucial factor in their midterm vote for Congress.
1
1
7
288
Lol Anthropic is fully capable of smoking all GPT 6.2 models right now. They’ve done 4 models already that surpass the Fable capability class
Image from Jesse Pan (@thefusioncube)
METR (@METR_Evals) profile pictureMETR
@METR_EvalsNov 5
Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development. We are now publishing a report summarizing the results of this risk assessment.
5
307
On days like today, ask yourself: What fundamentally changed in the last 24 hours? The catalyst is growing anxiety over the Democratic lawmakers called for a "ban on AI/SI", not just the Midterm Election. But an outright ban is extremely unrealistic. Buy the dip.
A market bloodbath, with all major tech stocks facing a major dip.

ALT A market bloodbath, with all major tech stocks facing a major dip.

Donald Trump, doubling down on his stance of not slowing down American AI development.

ALT Donald Trump, doubling down on his stance of not slowing down American AI development.

1
7
290
HUMAN SLOP IS WORSE THAN AI SLOP. Opus 5.5 could never. Imagine some hacker just spammed some letters as the password and got in. teapawt.net/IntCyberDigest/status/…
International Cyber Digest (@IntCyberDigest) profile pictureInternational Cyber Digest
@IntCyberDigestNov 5
The bug is like a joke. On Alpine Linux, type 257 to 1024 characters of anything and you get root. 💀
1
3
253
THE AI BUBBLE JUST POPPED 🚨🚨🚨 MORE THAN 2 TRILLION DOLLARS JUST GOT WIPED FROM U.S. STOCKS!
Image from Jesse Pan (@thefusioncube)
3
8
466
The Mistral AI student plan is surprisingly good, offering $150 in usage for only $6 per month! This can be used for models like Deepseek and GLM!
Image from Jesse Pan (@thefusioncube)
Image from Jesse Pan (@thefusioncube)
2
330
Lol content creators are now being paid to spread Anti-AI sentiment. This is directly threatening innovation and technological advancements which could benefit us all. teapawt.net/TaylorLorenz/status/21…
Taylor Lorenz (@TaylorLorenz) profile pictureTaylor Lorenz
@TaylorLorenzNov 4
I’m once again being offered a lot of money to promote anti-AI content
Image from Taylor Lorenz (@TaylorLorenz)
Image from Taylor Lorenz (@TaylorLorenz)
Image from Taylor Lorenz (@TaylorLorenz)
Image from Taylor Lorenz (@TaylorLorenz)
1
165
Local AI has its benefits, but would you really buy a 7k DGX Spark to run Deepseek/GLM 5.3 Flash just to avoid buying a $200 dollar plan for Claude Opus 5.5 or GPT 6.1 Sol? teapawt.net/MiaAI_lab/status/21182…
Mia (@MiaAI_lab) profile pictureMia
@MiaAI_labNov 5
Local AI will set you free. No usage limits. No price hikes. No model getting quietly worse overnight. You pick the model, run it & it stays the same. Your prompts never leave your machine. Own your intelligence. Fully private.
6
1
6
222
Load more