It took 234,760 tool calls and 6,852 session logs for Stella Laurenzo, AMD’s Director of AI, to prove what every Claude Code user has been screaming about. Her conclusion: Claude “cannot be trusted to perform complex engineering tasks.” 1,085 HN points. Anthropic’s own Claude Code creator showed up in the comments to confirm the bug.
It lands the same day Anthropic announced a 3.5 gigawatt compute deal that explains exactly why your token quota keeps melting.
In today’s indie hacker news:
- AMD’s AI Director torches Claude Code with 234K data points
- Anthropic locks in 4.5 GW of TPUs as revenue triples to $30B
- Indie dev ships MIT speech-to-text for macOS, no cloud
- Freestyle (YC S24) ships VMs that fork in 400ms with full memory
- Sam Altman New Yorker profile, fake AI singer, REvil bosses doxed
TOP STORIES
THE 234,000 TOOL CALL RECEIPT
AMD’s Director of AI proves Claude Code went lazy with 6,852 session logs
Stella Laurenzo runs AI at AMD. She filed a GitHub issue with 234,760 tool calls, 17,871 thinking blocks, and 6,852 session logs. Verdict: “Claude cannot be trusted to perform complex engineering tasks.” Read-to-edit ratio dropped from 6.6 to 2.0. Stop hook violations went from zero to 173 in 17 days. Thinking depth fell 67% in late February. And 33.7% of edits were made to files Claude had not read. One in three edits, blind.
The details:
- 80x more API requests in March (119,341) vs February (1,498) for the same human effort
- Root cause: a staged “redact-thinking-2026-02-12” rollout that hit 100% by March 12
- 191,000 lines merged in one February weekend, non-functional by mid-March
- Boris Cherny (Claude Code creator, @bcherny) confirmed on HN
- Workaround:
CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1andeffortLevel: high - 1,085 HN points, The Register coverage
Why builders care: If you run Claude Code on auto-accept, your agent is making 1 in 3 edits blind. Set effortLevel: high, drop ultrathink into hard prompts, roll back to v2.1.34 if token costs went sideways. Narrow scope on multi-agent fleets until Anthropic ships a real fix.
Work from any WiFi like it's your home network. NordVPN's Meshnet runs a free private mesh between your laptop, dev box, and home server. SSH from a café without exposing a port, the way you'd use Tailscale. The paid VPN on top lets you test geo-fenced Stripe checkouts or feature flags from any country.
We get a cut if you sign up. Only added for tools we use ourselves.
THE 4.5 GIGAWATT BACKSTORY
Anthropic locks in 3.5 GW of Google TPUs via Broadcom as revenue hits $30B run rate
Anthropic will access 3.5 gigawatts of Google TPU compute via Broadcom starting 2027, on top of the 1+ GW online in 2026. Total: 4.5+ GW. For scale, 1 GW powers Seattle. The revenue numbers bury the compute story: run rate jumped from $9B end of 2025 to over $30B in April 2026. 3.3x in four months. Enterprise customers spending $1M+ on Claude doubled from 500 to 1,000 in under two months. Top HN comment: “$19B to $30B annualized in a month? Feels like the lede is buried here.”
The details:
- $21B in Broadcom orders revealed in December 2025 ($10B Q3, $11B Q4)
- Covers nearly 1 million Google TPU v7p (Ironwood) units delivered as full racks by late 2026
- Mizuho estimates Broadcom AI revenue from Anthropic: $42B in 2027
- Multi-cloud: AWS Trainium, Google TPUs, Nvidia GPUs, Azure Foundry
- 2027 tranche contingent on “Anthropic’s continued commercial success” per Broadcom 8-K
Why builders care: Your Claude Code rate limits exist because demand outran compute. This deal is the fix, but relief lands 2027. Long term: cheaper inference, bigger context windows, Anthropic finally in Stargate territory.
LOCAL VOICE, ZERO SUBSCRIPTION
Ghost Pepper: hold-to-talk speech-to-text for macOS, MIT licensed, $0
Matt Hartman shipped Ghost Pepper. Hold Control, speak, release. Text auto-pastes wherever your cursor is. 100% on-device via WhisperKit plus a local Qwen 2.5 0.8B model that strips filler words. No cloud, no audio on disk, no subscription. The clever bit is the two-model architecture: Whisper transcribes, Qwen cleans up. Whole stack is about 1 GB on the Apple Neural Engine. Wispr Flow charges $144/year. SuperWhisper charges $250 lifetime. Ghost Pepper is MIT Swift you can fork.
The details:
- 1,200+ GitHub stars in roughly one day after Show HN
- Whisper small.en (~466 MB) default, Parakeet v3 (~1.4 GB) optional
- Apple Silicon only, macOS 14+, no Intel
- 4 install steps, no terminal required
- 371 HN points, the thread is “a support group for people who built the same app”
Why builders care: A working blueprint for a local voice layer you can drop into any macOS agent. The Whisper-plus-cleanup-LLM pattern transfers to anything you’re building. And if you just want to dictate code, this replaces Wispr Flow at zero cost.
THE SANDBOX THAT FORKS LIKE A BROWSER TAB
Freestyle (YC S24) ships full Linux VMs that fork in milliseconds with all in-memory state
Freestyle launched on HN with full Linux VMs (not containers, not Firecracker microVMs) that boot in 700ms and live-fork in milliseconds. Founder Ben Swerdlow: “We’re forking the whole memory of it. If you’re half way down a browser page with animations running, they’ll be in the same place in all the forks.” The pitch: coding agents need to checkpoint a working state, branch into 10 variants, and discard the losers. Containers can’t do that. Founders Ben Swerdlow and Jacob Zwang met in 10th-grade calculus, both became Apple Senior Engineers, both dropped out.
The details:
- 700ms cold start via memory snapshot restoration
- Nested virtualization: Docker, KVM, eBPF, FUSE, systemd inside VMs
- Free tier: 10 concurrent VMs, 500 runs/month, no credit card
- Pro: $500/month for 400 concurrent VMs
- Customers: Onlook, Wordware, HeroUI, vly.ai, Rork, Vibeflow
- 269 HN points, 146 comments
Why builders care: If you’re building any coding agent, your sandbox provider becomes critical infrastructure. Live forking is the missing primitive for branch-and-discard loops: checkpoint, fan out 10 hypotheses, pick the winner, kill the rest. At $500/month for 400 concurrent VMs, the math finally works.
TRENDING TODAY
Sam Altman New Yorker profile (1,385 HN points). Ronan Farrow cites internal memos from Sutskever and Amodei. A former OpenAI board member: “He’s unconstrained by truth.” If you build on OpenAI, multi-provider strategy just became risk management.
Fake AI singer takes 11 iTunes spots. One creator, AI tools, 11 chart positions. Luminate confirmed only 6,900 actual sales. The arbitrage isn’t “AI music is good.” It’s “iTunes weighting rewards download bursts.” Same broken weighting on KDP, App Store, Etsy, Substack.
Germany doxes REvil and GandCrab bosses. BKA named Daniil Shchukin, 31, as “UNKN.” 130+ acts of computer sabotage 2019-2021. SaaS founder takeaway: encrypt backups offline, segment customer data, treat ransomware as near-certain.
STACK OF THE DAY
Datakool - Privacy-first Google Analytics replacement. No cookies, sub-1KB script, no GDPR consent banner. Tracks page views, referrers, devices, custom events without fingerprinting. Free tier covers small sites. Drop-in script, 5-minute setup.
Not sponsored. We just feature tools builders would actually use.
BOOKMARKED TODAY
Quantum computing timelines (465 HN points). Filippo Valsorda explains why CRQC is decades out. Read before you panic-rotate to post-quantum crypto.
What being ripped off taught me (408 HN points). A founder on getting cloned and undercut. Distribution moats beat product moats every time.
Every GPU That Mattered. Interactive data viz of every consumer GPU since 1999. Plot raw transistors against price against year. Addictive.
Curated by AI, built by a human. Get this daily: indiehacker.news | X | Telegram