Sandeep Alluru

[Sandeep Alluru @sandeep_alluru]

@kimmonismus 4 studios over thunderbolt 5 is apple's 3x inference claim vs one box. that's the cluster they added this week.

[Sandeep Alluru @sandeep_alluru]

@TheTuringPost 3 lines out for every 1 they added. that's the haystack 3.0 diff from july 20. 2.31 still gets security patches through october.

[Sandeep Alluru @sandeep_alluru]

@Zai_org $10B over any 12 months, and only if you're renting it out as an api. running it on your own box is just the license.

[Sandeep Alluru @sandeep_alluru]

An AI agent recovered a quantum laser lock 99.3% of the time — no human.

Anthropic opened MHS — a preview so agents can run lab hardware.

> Setup: weeks/months → hours/minutes
> CMU dose-response ~3× faster
> QuEra lock recovery 99.3%

Would you let Claude run the bench?

[Sandeep Alluru @sandeep_alluru]

@ArtificialAnlys @perplexity_ai five bucks a thousand searches, same for low, medium, and high. so bumping context isn't a search-price thing.

[Sandeep Alluru @sandeep_alluru]

@GoogleDeepMind @FlowbyGoogle 40 seconds is the new cap. 10-second steps, and it now reads the last 10s of the clip you already made, not just the last second.

[Sandeep Alluru @sandeep_alluru]

@AnthropicAI 99.3% of the time it grabs the laser lock back with nobody there. quera, after they let an agent run the controller.

[Sandeep Alluru @sandeep_alluru]

Claude closed 85% of a deception gap. Six researchers closed 20% — same rules.

Anthropic had Claude search, train, and test 10 alignment failures.

> Held on models 4.7× larger
> Sonnet 5 → early Opus 4.8 in 60h · ~2k · ~15,000×

Would you let Claude align the next Claude?

[Sandeep Alluru @sandeep_alluru]

@Thom_Wolf 2506 ducks if that million is just the $399 robot. fewer if the extra packs got mixed in.

[Sandeep Alluru @sandeep_alluru]

@claudeai $50k extra credits if a project needs more than the seats. you apply. 1 to 25 people per group.

[Sandeep Alluru @sandeep_alluru]

@GoogleDeepMind 2.6% of words come out wrong on recordings, 4.0% when it's live. google says the final text shows up 70% faster than chirp 3.

[Sandeep Alluru @sandeep_alluru]

A 770B model just went Apache 2.0 — with 1M context.

Tencent shipped Hy4 preview: 49B-active MoE, FP8 too.

> 770B total · 49B active · 1M ctx
> Vendor SWE-Pro 65.7 · DeepSWE 64.3
> Tencent blind eval 2.99 vs GLM 2.92 / Kimi 2.94

Downloading?

[Sandeep Alluru @sandeep_alluru]

@lmsysorg @Alibaba_Qwen @NVIDIAAI @AIatAMD 3.3 extra tokens land each time it guesses ahead. that's sitting under the 540 tok/s.

[Sandeep Alluru @sandeep_alluru]

@Reuters $150M annualized is the number under $12.9B. last round was $4.5B in 2023.

[Sandeep Alluru @sandeep_alluru]

@business 23.2T tokens this week on openrouter. they said the weights drop tonight.

[Sandeep Alluru @sandeep_alluru]

The OpenRouter stealth #1 is now a MIT download — and a 57.

https://t.co/zBNCplx9D4 shipped GLM-5.3-Flash (ox-alpha): 320B, 18B active.

> AA Index 57 · $0.045/task discounted
> DeepSWE 63.4 vs GLM-5.2 46.2 (vendor)
> AutomationBench 48.8 vs 26.2 (vendor)

Pulling the weights?

[Sandeep Alluru @sandeep_alluru]

@Alibaba_Qwen 8.6x the prompt throughput of 3.7-plus at 1M when 90% is already cached. that's the serving number under the 6B-active line.

[Sandeep Alluru @sandeep_alluru]

@OpenAI 1459 tok/s per user on gpt-oss 120B. next word in 0.69 ms.

[Sandeep Alluru @sandeep_alluru]

@arena @Alibaba_Qwen 52 on artificial analysis for the same 27B. it's #6 on consumer product (1587) and #8 on gaming (1645) too.

[Sandeep Alluru @sandeep_alluru]

The free mystery model that doubled DeepSeek's usage is a GLM.

https://t.co/vwpdYaIxxe confirmed Ox Alpha to Bloomberg. Weights drop tonight.

> #1 OpenRouter usage
> 2× DeepSeek · still $0
> New GLM · weights tonight

Pulling the checkpoint?