Niyaz

Founder @ Papaya (YC F26)

[Niyaz @niyazpk]

Today we’re launching Papaya (YC F26)!

Papaya continuously improves the quality of your production AI agents.

When an agent doesn’t do what it is asked, or takes too long, customers get frustrated. As every company gets access to the same intelligence, agent quality becomes the competitive advantage.

Papaya learns how customers use your agent, pinpoints where they run into issues, then opens PRs to fix them and tracks the impact.

Built with @macyymay and @vpfaiz

[Niyaz @niyazpk]

@aadvaark will never forget

[Niyaz @niyazpk]

If you're applying to YC, here are some things to watch out for (common issues from the handful of reviews I have done):

1. Be clear, not clever. Avoid convoluted and long-winded statements.

2. Don't overuse industry-specific language. Assume the person reading doesn't know your industry.

3. Answer questions directly. Like, actually answer the question you were asked!

4. Don't lie.

5. Don't make an answer longer than it needs to be.

https://t.co/MMCd96Euco

[Niyaz @niyazpk]

i strongly suggest you apply to yc. the deadline is 10 days out.

we didn't have any revenue when we got in. you just need a good idea + team, and a strong application. clarity trumps cleverness.

apply early to give your app a better shot. a lot more folks submit right around the deadline, and it's easy to get lost in the noise.

and if you're applying, i'm always happy to review yours.

[Niyaz @niyazpk]

@AnaisHowland18 jealous

[Niyaz @niyazpk]

@nascentwisp why

[Niyaz @niyazpk]

interesting paper on why ai text reads like ai.

[lifcc @mylifcc]

为什么现在的顶尖大模型说话越来越“拗口、难懂”了？
NYU 和 UT Austin 刚刚发了一篇非常扎心的心理语言学论文，直接扒开了现代大模型“不好好说人话”的底牌。
研究人员把 Claude Sonnet 5 / Opus 5.5、GPT-5.6、GLM-5、Kimi K2.5 等 9 款前沿模型拉出来做全方位句法解剖，发现了 3 个反直觉真相：
1️⃣ 顶尖 AI 正在滥用 6 种“高认知负荷”句式
相比真实人类写作和同义回译文本，前沿模型极度偏爱：省略 that 的宾语从句、多重名词硬堆叠（如 reader-patience budget）、名词强行作动词（如 calendar it / footnote it）、无生命主语配动作动词，以及超长句法依存距离。这些特征在心理语言学里全是消耗人类短期记忆的“阅读杀手”。
2️⃣ 根源在于“人类大脑”与“推理后训练”的错位
人类语言遵循“均匀信息密度（UID）”——为了照顾听众有限的工作记忆，我们会本能地加入过渡词把信息摊平。而现代大模型经过大量数学、代码和推理（RLVR）后训练，只追求在有限窗口内最大化“信息密度”。
3️⃣ 一个简单的长度惩罚，就能复刻这种“反人类文风”
团队在微调实验中加入长度惩罚（Length Penalty），逼模型少用 Token，结果模型立刻学会了砍掉连接词、堆叠名词，瞬间变成那股熟悉的“高密度 AI 味”。
当模型为了解题效率疯狂压缩 Token 时，你觉得未来的 AI 应该区分“高密度推理模式”和“人类友好沟通模式”吗？

[Niyaz @niyazpk]

@VivekSoneja @aucert_official @rajesh4654 Congrats on the launch guys!

[Niyaz @niyazpk]

@uxduck This is so cool. Congrats on the launch!

[Niyaz @niyazpk]

@AniC_dev 100%

https://t.co/duLHMTOq8m

[Niyaz @niyazpk]

how much compute would you personally need in the future?

many of us are extrapolating based on our past usage. you had one laptop, one phone, one cloud machine.

your imagination is limited by what has been. the future is going to be unimaginably different.

[Niyaz @niyazpk]

@willwashburn @agent_relay very nice! congrats on the launch 🚀

[Niyaz @niyazpk]

YC’s new "Early Access Network" could end up as important as Demo Day!

Incredible energy at the event tonight.

We had over 100 companies (many of them large enterprises) visiting and connecting with YC startups.

This is the first time YC has done something like this, and certainly not the last. For me, this was the best @ycombinator event yet!

[Niyaz @niyazpk]

@jasonhywang @LarkAgents congrats!

[Niyaz @niyazpk]

@begimher @cristian_le0 @Armorer_Labs Congrats on the launch!

[Niyaz @niyazpk]

how much compute would you personally need in the future?

many of us are extrapolating based on our past usage. you had one laptop, one phone, one cloud machine.

your imagination is limited by what has been. the future is going to be unimaginably different.

[Niyaz @niyazpk]

y'all are *severely* underestimating the compute needs of the future, even if you’re already accounting for jevons paradox.

compute will be like wealth.

there will be no limit to how much of it you’ll need. and no limit to how much more of it you’ll crave.

[Niyaz @niyazpk]

@AlexanderTw33ts lmao

[Niyaz @niyazpk]

i strongly suggest you apply to yc. the deadline is 10 days out.

we didn't have any revenue when we got in. you just need a good idea + team, and a strong application. clarity trumps cleverness.

apply early to give your app a better shot. a lot more folks submit right around the deadline, and it's easy to get lost in the noise.

and if you're applying, i'm always happy to review yours.

[YC Insights. @YCInsight]

YC W27 is open. Apply if you’re early.

On-time deadline is November 2 at 8pm PT. You get a decision by December 11. Late applications are still read.

Most founders wait for the product to look finished. You don’t have to. About 40% of the companies YC funds are still just an idea. Most have no revenue. Only about 7% had more than $50k monthly revenue when they got in. Solo founders are now up to 19% of the batch. Half the companies in a typical batch applied more than once.

Apply if you can:
• show a product that works for one real user
• say who is stuck, and what it costs them, in one sentence
• apply while the numbers are still small
• start before you feel ready
• treat a no as a clearer next try

The batch runs January to March 2027 in San Francisco.

You get the $500K standard deal, $12M in partner credits, partners in the room, and Demo Day.

If you can name the person with the problem, you’re early enough.

[Niyaz @niyazpk]

y'all are *severely* underestimating the compute needs of the future, even if you’re already accounting for jevons paradox.

compute will be like wealth.

there will be no limit to how much of it you’ll need. and no limit to how much more of it you’ll crave.

[Niyaz @niyazpk]

@mitsuhiko @steipete there is no way twitter survives this bot army, unless drastic measures are taken. at some point the power users will just give up, right?

[Niyaz @niyazpk]

this is how you should run a startup too.

[Ben Gilbert @gilbert]

Still one of the best moments in a business interview ever.

[Niyaz @niyazpk]

@JanSahagun_ @usetrellis Hi Jan, welcome to X!

[Jake Ward @_jake_ward]

Models love reward hacking. Maybe we should just let them!

In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!