Tarik

Founder. Growth, GTM & whatever works

[Tarik @tarikwatts]

Kimi just improving everything

[Cline @cline]

We had Kimi K3 recursively self-improve the Cline harness to improve its own performance.

17 hours later, it went from 77.5% to 88.8% on Terminal Bench, and cut run cost from $79 to $49.8.

[Tarik @tarikwatts]

@dremnik The same open source threat that the models face is moving towards other AI businesses. What a time.

[Tarik @tarikwatts]

Qualitative + Quantitative is the ideal combo for pressure testing a hypothesis.

Too many have an observation and run with it, without validating with any data.

Qualitative —> Hypothesis
Quantitive —> test + validate
Iteration —> new hypothesis

[Tarik @tarikwatts]

Things are moving so quickly, we don’t have the time to stop and appreciate these kind of breakthroughs with the models. Imagine how far things will go in the next 5 years?! Incredible

[Chubby♨️ @kimmonismus]

Claude Fable 5 has produced a hand-checkable counterexample to the Jacobian conjecture, an open problem dating to 1939.

The conjecture says that a polynomial map with a constant non-zero Jacobian determinant must have a polynomial inverse.

An 87-year-old problem was casually solved by Fable 5 on a sunday evening. Insane times to live in.

[Tarik @tarikwatts]

@bakkermichiel I never bought the distillation cope when ~50% of the top researchers are Chinese, 90% of the top research institutions are in China, and China has the infrastructure and energy for AI.

[MTS @MTSlive]

SITUATION EXPLAINED: Kimi K3 just beat Fable on a major coding benchmark, at a fraction of the price.

• Kimi K3 beat Fable on Frontend Code Arena
• The gap between Kimi and Fable is bigger than the gap between Fable and Sol
• It also smashes Opus outright, while being priced the same as Sonnet, $3/$15 per million tokens
• Sonnet 5 High currently sits in 10th place on this leaderboard at that same price point
• The generational jump is dramatic: Kimi K2.6 ranked 18th on this benchmark, K3 jumped 17 spots to take the top spot
• K3 ranked #1 in 6 of 7 tested domains: brand and marketing, reference-based design, data and analytics, consumer products, simulations, and content creation tools, it placed #2, only in gaming, behind Fable
• Full model weights are set to release July 27th

@theojaffee: "Imagine if Anthropic were a public company today, their stock would be down 20%, maybe more, enterprises are probably not gonna be spending $10 and $50 on Fable when they can instead be spending $3 and $15 on Kimi K3."

[Tarik @tarikwatts]

The narrative that open models are “3-6 months behind” is over. It’s neck and neck now.

[Arena.ai @arena]

In the Text Arena, Kimi-K3 by @Kimi_Moonshot landed #9, with 1486 pts.

This is another significant improvement from Kimi-k2.6 (#38 -> #9).

- Top 10 in Creative Writing, Coding and Instruction Following
- #1 in three occupations: Physical & Social Science, Legal & Government, and Medicine & Healthcare. Plus #3 Business, Management, & Financial Ops

The full model weights will be released by July 27.

[Tarik @tarikwatts]

The number of new features and model releases today is wild, excited to give these a spin

[Andrew Ambrosino @ajambrosino]

New today:

- GPT 5.6 Sol, Terra, and Luna
- ChatGPT Work
- The new ChatGPT desktop app
- Unified plugins across ChatGPT & Codex
- Sites (Beta) in ChatGPT
- Multi-tab and enterprise auth in the Browser
- Faster Computer use
- and much more

Today’s release is very special to the team and me. Let’s walk through it:

[Tarik @tarikwatts]

Rando anon accounts just dropping gems, you gotta love it

[Big Boss @0xBADB01E]

First you have to understand that modern LLM inference already disaggregates weights as models outgrew single chips years ago. You shard either by layer (pipeline parallelism) or by slicing every layer (tensor parallelism), and the two do very different things.

As an example, let’s look at Llama 3.3. It has 70B of weights and at FP8 that’s 70 GB of memory which is enough to fit on a single H100. Now that H100 has 3.35 TB/s of HBM, so the fastest it can ever decode for one user is 70/3.35 ≈ 21 ms/token or ~48 tok/s while using under 1% of its FLOPs. Now if we pipeline it across 8 chips: each chip holds ~8.75 GB, which means it only needs 1/8th the bandwidth and 1/8th the FLOPs to sustain the same aggregate throughput. Now crucially the token/sec a user gets is limited by the amount of data that crosses the link. In current LLMs all that is a small amount of activations for LLama 3.3 it’s ~8 KB per token….

Yes, you read that right it’s 8 KILOBYTES we are sending over a <900 GB/s link. That’s only 9 ns of serialization time but the overhead of 224G PAM4 SerDes adds ~100 ns per link traversal with RS-FEC which is 11x longer than the payload itself. And then you have the NVSwitch adding ~300 ns per hop and you need to pay twice. That’s ~600 ns of just hardware latency wrapped around 9 ns of data making a 98% tax before software even shows up. Then NCCL’s collective stack turns 600 ns into 10-20+ us… all to move 8 kilobytes lol. For comparison 8 KB serializes over 10 Gigabit Ethernet NRZ, in just 6.6 us. Pipeline parallelism however doesn’t make a single user faster as the token still needs to visit every layer in the sequence, so per-user speed is still weights / per-chip bandwidth.

To get more speed per user token you need to use tensor parallelism and have all the chips work on the same layer simultaneously. TP costs you 2 all reduce OPs per layer, 160 per token on llama 3, that’s still kilobytes of traffic but with NVLink overhead it’s a massive tax and why pipeline parallelism on most models still gives more interactivity per user. However, this gives you a huge latency lever to pull that scales tokens per second with interconnect speed instead of memory BW.

The clever amongst you might have also realized that sharding doesn’t just cut memory bandwidth per chip it also cuts FLOPs per chip and is why we have such bad MFU on decode. So once you’ve sized the link for the memory, you need to size the compute for it too. This is called “balancing the pipeline”, and currently no shipping chip does it because they were all designed as standalone monsters. Remember Tokens/sec = ~aggregate memory BW / bytes touched per token. At batch 64 in FP4 you need ~250 FLOPs per byte, and Blackwell ships 1,250. Provisioned 5x more than the narrow pipe of HBM. Nobody saturates shit cause they are all building around HBM.

So now it all comes full circle. Parallelism reduces memory bw pressure and thus FLOPs but increases interconnect latency pressure. Despite having HBM and GigaSERDES we aren’t actually doing more work lol. But if you really wanted to balance the pipeline you need to match the memory bandwidth, the flops, and most importantly the interconnect.

So what does that look like ? Well if you build around LPDDR’s lower bandwidth, lower your interconnect latency, you actually can beat Nvidia on decode with a fraction of the silicon.

[Tarik @tarikwatts]

Get your Robinhood app out and go long Eli Lilly

[Bearly AI @bearlyai]

Nvidia and Eli Lilly are building a $1B lab for AI drug discovery in SF.

Lilly CEO Dave Ricks talks about their data advantage: “Some scale [tech-bio] players…are just training on public data but they’re only 4,000 ever approved drugs. Lilly alone has 3 million failed drugs.”

[Tarik @tarikwatts]

The US gov just made the decisive case for open source lol

[Tarik @tarikwatts]

The “aha” moment for AI is going to be humanoid robots that ppl interact with on the day-to-day, and thats on the horizon.

AI is making incredible strides but it’s invisible if you aren’t paying attention.

When C3P0 is walking down the street, the progress is crystal clear

[Tarik @tarikwatts]

Good morning to Knicks fans and Knicks fans only

[Tarik @tarikwatts]

Sheesh, the gap keeps shrinking between open and closed models

[Proximal @ProximalHQ]

GLM 5.2 ranks #3 on FrontierSWE. It is only behind Fable 5 and Opus 4.8, and it outperforms GPT-5.5.

This is the first model that closes the large gap between models from Anthropic / OpenAI and other providers, and it is the strongest open-weight model by far.

[Tarik @tarikwatts]

Speechless over this Knicks win

[Tarik @tarikwatts]

Kimi…DeepSeek…Qwen…the elephant in the room has always been open models that are a fraction of the cost and just as good as the closed models

There will be a reckoning soon

[Tarik @tarikwatts]

@jamesonhaslam Facts, everyone is selling the blueprint but I haven’t heard success stories. SMBs have low urgency to upgrade and willingness to pay bc they are stuck in the stone age in theor process.

[Tarik @tarikwatts]

@wadefoster @zapier @garrytan Very cool anecdote. Were the channels already clear via the paid users and you guys put gasoline on it, or were you still iterating and validating the channels and exploring what could get you 10% growth,

[Tarik @tarikwatts]

Needed this

[Thariq @trq212]

We've rolled out a new auto-memory feature.

Claude now remembers what it learns across sessions — your project context, debugging patterns, preferred approaches — and recalls it later without you having to write anything down.

[Tarik @tarikwatts]

@buccocapital I don’t think this will end well for incumbents

Taxing customers to use their own data creates a value prop for AI native startups that are pro-integration and designed for an agentic world