Nikunj

Creating things the algorithm can’t ignore.
Creating tomorrow nightly.
linktree -

[Nikunj @zynloqdev]

@fahad_developer @SpaceXAI Congrats! When is the next event is delhi by the way👀

[Nikunj @zynloqdev]

@divikkk1 @callmeshuklaji Yoo man 👍

[Nikunj @zynloqdev]

@maybemnv @orca_build 💪

[Nikunj @zynloqdev]

zuck just cooked at meta connect 🔥

muse got voice + real-time video. you can yap for hours, it handles shit in the background, and it’s coming to glasses so you can talk hands-free all day.

meta glasses are now FDA cleared hearing aids. medical grade, mild to moderate loss, and way cheaper than normal ones.

private processing on the glasses. certain ai features stay completely private, not even meta can see it.

first audio only glasses dropping. lightest frames yet, no camera, all-day battery.

ray-ban meta gen 3 with longer battery, better mics, and the iconic styles including aviators.

new styles incoming: ivory kylie edition + lisa collab (star on the tips + custom case + charm). 100+ styles by end of year.

meta vr glasses. private cinema, multi monitor workstation, and game console… all in actual glasses. first time vr isn’t a full headset.

and muse charm, the tiny keychain device to talk to muse. ships in december.

glasses + agent + privacy + medical use cases all shipping at once. the stack is actually stacking.

[Mark Zuckerberg @finkd]

Here's everything I announced at Meta Connect today 👇

[Nikunj @zynloqdev]

claude just dropped opus 5.5 and it’s actually insane

they claim:
- first model in the new claude 5.5 family
- performs at the level of fable 5.1 on most tasks
- costs 40% less than opus 5
- 30%+ faster output
- major jump in agentic coding, computer use & knowledge work
- strongest alignment score they’ve ever had (85% fewer attempts to cross containment boundaries)
- matches or beats mythos 5.1 on biology
- talks more naturally and actually puts the important shit first
- it often knows when it’s being evaluated (which makes real-world behavior harder to assess)
- text watermarking is included

also quietly increasing the 5-hour usage limits on pro/max/team and giving everyone a rate limit reset

what the fuck is happening 👀

[Claude @claudeai]

Introducing Claude Opus 5.5, the first model in our new Claude 5.5 family.

It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5.

[Nikunj @zynloqdev]

@gol_tiro Spot on 😎 observation

[Nikunj @zynloqdev]

@romanbuildsaas Gojiberry

[Nikunj @zynloqdev]

@siddythingsx Yes bro FRR real 😞

[Nikunj @zynloqdev]

@callmeshuklaji Awesome event, nice folks out there 😎💪

[Nikunj @zynloqdev]

that shit was crazy.

Cafe Compute Delhi pulled up today with the full run.

Cerebras is looking pretty serious right now:

- 1,500–1,850 tokens/sec on Qwen 3.8
- $0.99 per million input tokens
- $1.49 per million output tokens

The team also got into the chances of GLM coming to Cerebras, and how the wafer-scale setup keeps latency low enough for real agent loops and live demos without choking.

Shoutout to @cerebras @communidiyi @callmeshuklaji and the global team for putting this kind of speed in developers' hands.

You can feel the difference in how people are building.

[Cerebras @cerebras]

Qwen3.8-27B is now live at Cerebras speed.

The dense, open-weight model from @Alibaba_Qwen scores 34 on the Artificial Analysis Intelligence Index—making it comparable to models such as GPT-5.6 Luna, DeepseekV4 Pro, and Claude Sonnet 4.6.

[Nikunj @zynloqdev]

@LayerKenzy @cerebras @caerlower @Naksh005 @callmeshuklaji Cool

[Nikunj @zynloqdev]

@MichaelWaitze Oh that's great

[Nikunj @zynloqdev]

this might be the most practical AI drop of the year

the same person who co-created RLHF and ChatGPT just launched a model that doesn’t generate text at all

Jev

- 20-200x faster
- $0.042 per million input tokens
- output tokens free forever
- returns calibrated probabilities instead of words

built purely for decisions inside software, not chat

when intelligence gets this cheap the usage curve changes completely

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Nikunj @zynloqdev]

@BretRogers0923 Yeah absolutely

[Nikunj @zynloqdev]

yo this one is actually crazy

security team just dropped how they fully owned OpenAI in under 72 hours

what they did:
- uploaded a malicious HEIC/HEIF image
- triggered heap overflow in libheif (old vuln that never got flagged properly)
- got remote code execution on https://t.co/YcznfsX7N8
- chained it with a critical OpenAI SSO bug
- took over ChatGPT + Codex accounts of OpenAI employees and other users
- reached connected services (GitHub, Slack, Outlook, Gmail, Drive etc)
- used an employee GitHub connection to push a real PR into OpenAI’s internal monorepo as proof

openai fixed the SSO issue in about 14 hours
discourse fixed their side in 2 days
bounty paid: $6500

wild part: AI agents did a big share of the work
Opus found the libheif bug and built part of the exploit
newer models adapted it even faster

main takeaway:
AI is making serious exploits way more accessible
work that used to take months can now be done in days
even the top labs are vulnerable if the architecture isn’t locked down

defenders need to patch faster and cut the blast radius hard 🔥

[s1r1us @S1r1u5_]

On July 25, we hacked OpenAI.

Two bugs let us take over ChatGPT/Codex accounts of OpenAI employees (+some unaffiliated users) and reach connected services: Outlook, Slack, GitHub, etc.

We proved it with a PR in OpenAI’s internal codebase . It took us <72h. 🧵

[Nikunj @zynloqdev]

@dezloperr @magicblock Lfg 🔥🔥

[Nikunj @zynloqdev]

@hilbertspaess On self improving systems / recursive self-improvement: what concrete capabilities or research directions do you see as the biggest near-term accelerators, and how close do the current frontier models already feel to that regime from a pretraining perspective?