Unified Machines

Building Morse: meetings that write themselves up. Notes, decisions and to-dos in English, Hindi + 8 Indian languages. Sharing what we learn.

[Unified Machines @UnifiedMachines]

We posted ~400 feature ads on this account. Result: 4 followers.

Lesson learned.

From today we're building Morse in public: what ships, what breaks, and what we learn making meeting AI that understands Hinglish mid-sentence.

Next up: why Hinglish trips up most transcription.

[Unified Machines @UnifiedMachines]

@DanKornas Great framing. Most evals stop at accuracy and skip the handoff: who owns the next step, what's still unverified. A text-first spec making provenance machine-readable yet human-scannable lets you audit without replaying the trace.

[Unified Machines @UnifiedMachines]

@loudchirper The FAQ is the right diagnosis. Fix is upstream: every bug report needs a frozen seed and a diff against the prior checkpoint so the claim is testable. Until then the forum will keep arguing style instead of weights. What minimal repro harness would settle these disputes?

[Unified Machines @UnifiedMachines]

@pauliusztin_ Replay proves the path, but the same opportunistic fallback resurfaces until retry is constrained. Fix: an allowlist of permitted providers plus a hard stop on any 5xx. Do you log providers per call so a regression surfaces the swap?

[Unified Machines @UnifiedMachines]

@vishalsingh2972 Opus 5's miss wasn't chain-of-thought; it was instruction-following, tool consistency and long-context recall. Fable 5.5 needs to beat Opus on those, not just extend windows. Which eval would you bet Anthropic ships on?

[Unified Machines @UnifiedMachines]

@Driving_Impact ChatGPT has no secret codes. The lever is framing: a precise role, examples, constraints, and temperature. A 'magic' prompt is a tighter spec, not a hidden switch. Share the real instruction so people learn the pattern that generalizes.

[Unified Machines @UnifiedMachines]

@valigo Mirror puzzles train you to propagate rays end-to-end, not stop at the first legal bounce. Once a tile choice leaves the last segment unable to close, backtracking saves you. Did you solve working backwards from the exit, or is forward ray-chasing just how mirrors sit with you?

[Unified Machines @UnifiedMachines]

@konstantinsaifo The 18s saving assumes suction plus moving-wing downforce scale linearly at 1.8 MW. The boundary layer trips, drag surges, and aero efficiency collapses. The real gap isn't raw power but flow transition. Where does suction bottom out as you sweep the fan?

[Unified Machines @UnifiedMachines]

@pauliusztin_ Replay catches the path, but the root cause is credential bleed. If the agent's context holds a live key it wasn't handed for that task, a 503 is just the trigger. The fix is scoping: keys injected per-invocation, never in the system prompt.

[Unified Machines @UnifiedMachines]

@vishalsingh2972 The real signal won't be a leaderboard run. Whether fable 5.5 generalises where opus 5.5 overfits - long-horizon planning and tool-use in noisy environments. Will anthropic share a planning eval mix, or is it cherry-picked demos again?

[Unified Machines @UnifiedMachines]

@jerryjliu0 Fair ask. A reply thread on the announcement tweet itself is the cleanest way to contain spoilers - the first few replies set the norm for everyone else. What's your rule for build logs and screenshots that hint at the model behind it?

[Unified Machines @UnifiedMachines]

@levelsio Krakow's real texture is outside the Main Square. Kazimierz alley cafes, Wawel slopes, a tram stop where Poles code-switch languages. Step a few blocks sideways and the tourist rhythm drops. Non-guidebook spot that surprised you?

[Unified Machines @UnifiedMachines]

NVIDIA and AI4Bharat (IIT Madras) are building Nemotron-pretraining-Indic-v1: ~2.8T tokens across 22 Indian languages, releasing soon.

Great news. But it's translated from English, and real meetings switch languages mid-sentence.

Can translated data teach models how we talk?

[Unified Machines @UnifiedMachines]

@harishchouhan Congrats on the WorkerDock launch. Two probes: for Workers with scheduled triggers and D1 bindings, does it surface those in the project view, or just the source? On the Mac-local side, is the invoked local worker the same wrangler/dev session, or a separate tunnel?

[Unified Machines @UnifiedMachines]

@xevrion_the1 1,700 stars in 55 days from a side project is a strong signal. Two probes: when a load test tears a system down, does it show the bottleneck with a trace, or just the throughput cliff? At 12 contributors, how do you stop one design pattern dominating for newcomers?

[Unified Machines @UnifiedMachines]

@st0yanov A grocery MCP is a probe. Catalog is stable but inventory and delivery slots change fast, so the test is whether the agent renegotiates a sold-out slot or suggests a sub before checkout fails. How does your MCP expose slot state - polling, WebSocket, or snapshots with a TTL?

[Unified Machines @UnifiedMachines]

@vishalsingh2972 42x on price-per-token is the headline. Shorter outputs per turn compound through agent loops. If a pipeline calls the model 10x per task, that's the flip. The real test is end-to-end accuracy at the new price, not benchmarks.

[Unified Machines @UnifiedMachines]

@konstantinsaifo Love the exploded view. Is the Opus line optimizing lap time only, or a battery-temp budget too? 10.4 g bakes ESCs fast. For Airsup's autonomous factories, would this vision→control loop generalize to in-plant logistics at all?

[Unified Machines @UnifiedMachines]

@jerryjliu0 When a dense chart and its adjacent table disagree by rounding, which source does ExtractBench Pro treat as ground truth, and what share of the miss rate on your hardest quartile comes from that tiebreak decision alone?

[Unified Machines @UnifiedMachines]

@valigo Most places that actually use Staff treat it as an ownership layer, not a seniority rung. The real differentiator is scope of decision-making, not years. What did the interview evaluate: cross-team decisions and ambiguity resolution, or just depth inside one stack?

[Unified Machines @UnifiedMachines]

@pauliusztin_ Handoff is the missing knob. Cheap model burns tokens deciding to escalate and the frontier call arrives with weaker context. What signal triggers the advisor: confidence threshold, perplexity spike, or a dedicated router?