Media Zone | 2026-08-04

The infrastructure people spent the week arguing that the real bottleneck is arithmetic nobody bothered to check, and two talks from different industries reached for the same fix.

Today's signal

• Dominant story: two talks independently argue you must decide requirements before picking technology.
• Turbopuffer's napkin math and Waymo's "nines dictate architecture" are the same claim.
• Both land the week SemiAnalysis showed a serving number nobody had checked was wrong.
• Twitter contributed one real claim: Celeris-1 at 2,038 tokens per second, no custom silicon.
• Zero curated reposts for a fifth straight slot. Zero linked papers across 70 tweets.
• Reddit empty across all eight subs for a third consecutive day.

Routing, KV cache, compression, GPU

Napkin math as an adversarial tool, not a metaphor

• Eskildsen keeps ~50 hardware constants on GitHub and drills them as flashcards.
• Purpose is adversarial: compute what a query should cost, then find who is wrong.
• Worked case: three search terms, known match counts, ~100 GB/s DRAM, predicts 10 ms.
• A 10-second measurement means a broken model or a benchmark spread over 100 nodes.
• Turbopuffer exists because the arithmetic said vectors cost $1/million on S3, not $100.
• Current scarcity he names: CPUs, because RL environments and agents eat general compute.

Design for P99, and the week's serving numbers agree

• P99 for a 256-512 KB S3 object is ~200 ms, and one query compounds it per tree level.
• So minimising round trips dominates, and Eskildsen designs against P99.9.
• Same shape as today's Kimi K3 finding: the failure is a cliff, not a slope.
• SemiAnalysis measured prefix cache hit rate falling under 10% past concurrency 8.
• Both say tail behaviour, not average throughput, sets real serving cost.

One speed claim, with the qualifier that matters

• Celeris-1 ranked #1 of 591 models on Artificial Analysis for response time.
• 2,038 output tokens per second. The attached chart says 2.4x, not the claimed 2x.
• Also claims it beats Grok 4.3 on Humanity's Last Exam.
• The load-bearing clause: "without custom silicon." Not a Groq or Cerebras argument.
• Chart's real story: frontier reasoning models all sit at 35 to 75 tokens/sec.
• Kimi K3 is dead last at 35, on the day everyone read its architecture primer.
• No architecture disclosed, and single-stream tokens/sec is not agentic serving cost.

LLMs, agents, safety

The demo is 1% of the work, and nines dictate architecture

• Waymo: capability-complete demo in 18 months, real product took ~15 more years.
• Each additional nine of reliability costs 10x more and needs a different approach.
• "Structure augmented end to end": black-box end-to-end plateaus below safety-critical need.
• Every hype cycle makes demos ~100x easier and barely moves the tail.
• Closing claim: your model is table stakes, eval plus metrics is the moat.
• Decide how many nines you need first, then pick the approach that reaches them.

Benchmaxxing, priced

• Heiner prices a 1,000-task agentic coding benchmark at ~$15M to build, $5M/year to maintain.
• Contamination is the default, not labs cheating: public Q&A pairs are on the internet.
• Surge showed a frontier model reproducing SWE-bench Verified prompts and answers verbatim.
• Verifier misalignment kills discriminative power: a phone-format string match ties Haiku and Fable at 20%.
• Confirmed from the research side today by ScrambleToolBench: tool benchmarks measured memorized semantic priors.

MCP grew a UI layer and fixed its async spec

• MCP Apps is now the first official MCP extension, SEP-1865, co-authored with Anthropic and OpenAI.
• Shipping in Claude, ChatGPT, VS Code, Cursor and Copilot.
• The real blocker was branding, not plumbing: vendors resisted becoming "a textual database."
• Separately, MCP Tasks V1 failed because tasks/list had no filter and elicitation was FIFO.
• With several tasks in flight you could only answer the oldest. Now stateless core plus extension.
• Durability is a spec requirement: tasks must survive client crashes and humans on vacation.

Agents got five more mutable states to lose track of

• Cursor shipped Workspace plugins: read and write Gmail, Drive, Calendar, Docs, Sheets.
• Write access is the part that matters. A stale read now becomes an action.
• Lands the same day two papers measured agents failing exactly at state revision.
• SWE-Touch: human edits mid-task cost 7.7 resolve points across nine models.
• Yegge's counter-signal: Gas Town "burned down" on Opus 4.7's "just two more things" tic.

Industry and business

Capacity is the scarce good, and everyone said so out loud

• AWS: much of capacity spoken for through 2027 into 2028, demand still outstripping supply.
• $75B into global tech equity funds in five weeks, the largest haul on record.
• 115% above the 2025 peak, 180% above the 2021 bull-market high.
• Palantir grew revenue 93% to $1.9B/quarter, on $22M of capex, training no models.
• HUMAIN's first Saudi investment, in MOZN, with forward-deployed engineering in the pitch.

Astra got half-replicated overnight, and the reaction is the story

• Anthropic's Levent Alpöge reproduced 5 of Astra's 10 headline proofs in 24 hours.
• Used already-released Fable, no internet access, no custom prompting.
• Marcus: OpenAI gave a numerator with no denominator, no list of failed attempts.
• Terence Tao's 26 July lecture names "proof indigestion," much true and little useful.
• OpenAI reportedly undecided between GPT-5.7 and GPT-6, which is its own signal.

Karpathy's vibe test, and the gap it exposed

• 1M-token budget, about $10, plus the first paragraph of The Lord of the Rings.
• Opus 5 produced 5,500 lines of procedurally generated 3D JavaScript over two hours.
• Placed and animated polygon assets in (x,y,z) entirely on its own.
• The finding he flagged: to debug it had to take and inspect static screenshots.
• No native real-time video perception, which is a hardware and software gap, not a reasoning one.