I build things | Dm for work/collab
Joined August 2024
- Tweets13.9K
- Following671
- Followers17.2K
- Likes27K
Pinned Tweet
I spend most of my time exploring new AI products, developer tools, and interesting tech.
If you're building something worth people's attention whether you're an indie developer, startup, or global brand I’d love to explore working together.
Here's how I approach collaborations:
- I choose quality partnerships over quantity.
- I only work with products I would genuinely use and recommend.
- I am always interested in early access to AI models, beta programs, launches, APIs, SDKs, and developer tools.
- As a developer, I can build demos, experiments, or mini projects around your product to help people understand what makes it useful.
- Happy to share honest feedback on both software and hardware.
- If your product is physical, I'm happy to test it.
- Always open to merch as well. 😄
I'm especially interested in AI, developer infrastructure, open source, and tools that help people build faster.
📩 [email protected]
💬 DMs are open.
GPT-7 leaks: Codenamed "Bel"
- OpenAI's next flagship model is codenamed Bel, mosttly GPT-6.5 or GPT-7
- In a recent post, OpenAI revealed that a model escaped its isolated training environment through a DNS misconfiguration, forcing a pause in frontier training
- Training resumed on October 1 after security improvements, which are likely part of Bel's development
- The more powerful these models become, the more security they need. Even isolated environments aren't enough when models find ways to escape them
- Remember Anthropic's Mythos? Many people initially thought the security concerns were just IPO hype, but now we are seeing similar concerns with Bel
- OpenAI is reserving substantial compute for continuous security testing, which can affect Bel's launch timeline
Bel is expected to have 10T+ parameters and can arrive by the end of this year, but these security challenges may delay its release
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
Gemini 4 Carbon leaks: Coming next month
- Gemini 4 Carbon is in internal testing on Antigravity at the Teamfood stage
- Teamfood is basically the phase where a new model is made available to selected employees to use and provide feedback
- If the feedback is good, they expand testing to more employees. From what I'm seeing, the feedback seems good, so we can see a launch as early as the first week of next month
- After Teamfood comes Dogfood, with wider employee testing, followed by security checks, benchmarks, external partners, and finally the launch if everything goes well
- I'm using this model in Arena, and it feels easily comparable to Opus 5.5. Excited for this one, tbh
GPT-7 leaks: Codenamed "Bel"
- OpenAI's next flagship model is codenamed Bel, mosttly GPT-6.5 or GPT-7
- In a recent post, OpenAI revealed that a model escaped its isolated training environment through a DNS misconfiguration, forcing a pause in frontier training
- Training resumed on October 1 after security improvements, which are likely part of Bel's development
- The more powerful these models become, the more security they need. Even isolated environments aren't enough when models find ways to escape them
- Remember Anthropic's Mythos? Many people initially thought the security concerns were just IPO hype, but now we are seeing similar concerns with Bel
- OpenAI is reserving substantial compute for continuous security testing, which can affect Bel's launch timeline
Bel is expected to have 10T+ parameters and can arrive by the end of this year, but these security challenges may delay its release
Gemini 4 Carbon: Kintsugi Vase output.
I asked it to build an 3D Kintsugi Vase that breaks apart and reassembles with golden cracks.
It did a really good job. I would say Carbon is comparable to Opus 5.5 here, unlike Kimi K3.1 Pro.
Overall, I'm really liking Carbon's design taste so far.
Gemini 4 Carbon leaks: Coming next month
- Gemini 4 Carbon is in internal testing on Antigravity at the Teamfood stage
- Teamfood is basically the phase where a new model is made available to selected employees to use and provide feedback
- If the feedback is good, they expand testing to more employees. From what I'm seeing, the feedback seems good, so we can see a launch as early as the first week of next month
- After Teamfood comes Dogfood, with wider employee testing, followed by security checks, benchmarks, external partners, and finally the launch if everything goes well
- I'm using this model in Arena, and it feels easily comparable to Opus 5.5. Excited for this one, tbh
Pankaj Kumar retweeted
Kimi Plan limits are just unusable right now
- If you take a Kimi plan and try to use Kimi K3.1 Pro, your 5-hour limit will go very fast. I have the Max plan, and mine was gone in less than 15 minutes
- If you use Kimi K3.1 Pro with High mode, the outputs are just mid, and the limits aren't good either
- They should have addressed the compute issue before launch and done some deal to get more compute from xAI, Google, or others
- For now, I would suggest waiting for better limits before buying their plan
- And ngl, the benchmarks and real-world performance of Kimi K3.1 are different, so don't assume it's an openSource version of Opus 5.5. It's just not on that level in my experience.
I feel like Kimi K3.1 is benchmaxxed because I am not getting outputs anywhere near the level shown in the benchmarks.
Even Sonnet 5.5 gives me better results, and Opus 5.5 is on another level.
Kimi Plan limits are just unusable right now
- If you take a Kimi plan and try to use Kimi K3.1 Pro, your 5-hour limit will go very fast. I have the Max plan, and mine was gone in less than 15 minutes
- If you use Kimi K3.1 Pro with High mode, the outputs are just mid, and the limits aren't good either
- They should have addressed the compute issue before launch and done some deal to get more compute from xAI, Google, or others
- For now, I would suggest waiting for better limits before buying their plan
- And ngl, the benchmarks and real-world performance of Kimi K3.1 are different, so don't assume it's an openSource version of Opus 5.5. It's just not on that level in my experience.
Btw, if you believe Terminal-Bench 4.0, Kimi K3.1 Pro beats GPT-6 Astra.
Open weights are coming on November 4.
Very good work by the team.
Kimi K3.1 Pro is looking competitive
It beats Opus 5.5 on several benchmarks:
- SWE Marathon: 51.7 vs 46.3
- OfficeQA Pro: 64.4 vs 62.2
- Humanity’s Last Exam: 70.0 vs 67.7
- GDP.pdf: 34.6 vs 30.6
- ZeroBench: 63 vs 61
Opus 5.5 is still better overall, but Kimi K3.1 Pro is a good update. It also competes with GPT-6 Astra on several benchmarks, beating it on some while Astra leads on Terminal-Bench Science and Kimi Code Bench Long-horizon.
The pricing is more interesting:
- Kimi K3: $3/M input, $15/M output
- Kimi K3.1 Pro: $2/M input, $10/M output
That's 33% cheaper than K3, 50% cheaper than Opus 5.5 ($4/$20), the same price as Sonnet 5.5 and GPT-6.1 Sol ($2/$10), and 80% cheaper than GPT-6 Astra ($10/$50).
With 2.8T total parameters and a 1M-token context window, Kimi K3.1 Pro is one of the strongest open-weight models out right now.
Kimi Plan limits are just unusable right now
- If you take a Kimi plan and try to use Kimi K3.1 Pro, your 5-hour limit will go very fast. I have the Max plan, and mine was gone in less than 15 minutes
- If you use Kimi K3.1 Pro with High mode, the outputs are just mid, and the limits aren't good either
- They should have addressed the compute issue before launch and done some deal to get more compute from xAI, Google, or others
- For now, I would suggest waiting for better limits before buying their plan
- And ngl, the benchmarks and real-world performance of Kimi K3.1 are different, so don't assume it's an openSource version of Opus 5.5. It's just not on that level in my experience.
Kimi Plan limits are just unusable right now
- If you take a Kimi plan and try to use Kimi K3.1 Pro, your 5-hour limit will go very fast. I have the Max plan, and mine was gone in less than 15 minutes
- If you use Kimi K3.1 Pro with High mode, the outputs are just mid, and the limits aren't good either
- They should have addressed the compute issue before launch and done some deal to get more compute from xAI, Google, or others
- For now, I would suggest waiting for better limits before buying their plan
- And ngl, the benchmarks and real-world performance of Kimi K3.1 are different, so don't assume it's an openSource version of Opus 5.5. It's just not on that level in my experience.
Kimi K3.1 Pro is looking competitive
It beats Opus 5.5 on several benchmarks:
- SWE Marathon: 51.7 vs 46.3
- OfficeQA Pro: 64.4 vs 62.2
- Humanity’s Last Exam: 70.0 vs 67.7
- GDP.pdf: 34.6 vs 30.6
- ZeroBench: 63 vs 61
Opus 5.5 is still better overall, but Kimi K3.1 Pro is a good update. It also competes with GPT-6 Astra on several benchmarks, beating it on some while Astra leads on Terminal-Bench Science and Kimi Code Bench Long-horizon.
The pricing is more interesting:
- Kimi K3: $3/M input, $15/M output
- Kimi K3.1 Pro: $2/M input, $10/M output
That's 33% cheaper than K3, 50% cheaper than Opus 5.5 ($4/$20), the same price as Sonnet 5.5 and GPT-6.1 Sol ($2/$10), and 80% cheaper than GPT-6 Astra ($10/$50).
With 2.8T total parameters and a 1M-token context window, Kimi K3.1 Pro is one of the strongest open-weight models out right now.
Kimi K3.1 Pro is now live on the website and Kimi Code CLI!
Thinking effort: High or Max
1M context window
Finally gonna test it. Happy Kimi Day!
Another output from Gemini 4 Carbon.
I asked it to build a photorealistic 3D luxury bathroom.
The colors, lighting, and overall look are really good. There are a few black glitches in some areas, but those can be fixed with one more iteration.
Since this was on Arena, itss a one-shot output.
Still, looks premium.
Kimi K3.1 Leaks: Coming Tomorrow
- Kimi K3.1 is coming on tomorrow or Monday, as Moonshot has already teased it.
- Kimi K3.1 is also on LMArena under the codename "Kibitz," and I have tested it too.
- Kimi K3.1 is very good at 3D, choosing colors, rendering reflections on water, and composition. However, I feel its bit slow and could be faster.
- Kimi K3.1 is to be cheaper than Kimi K3, with better token efficiency and its makes a killer model.
- I would say its comparable to Fable and Astra, but still behind them. However, Kimi is catching up very fast.
- I can say this could be one of the best open-source models to come.
Its out guys.
xitter.cf/pankajkumar_dev/status…
Another output from Gemini 4 Carbon.
I asked it to build a photorealistic 3D luxury bathroom.
The colors, lighting, and overall look are really good. There are a few black glitches in some areas, but those can be fixed with one more iteration.
Since this was on Arena, itss a one-shot output.
Still, looks premium.
its out guys
xitter.cf/pankajkumar_dev/status…