Nithin Stephen Raphael

@byreachdesign | Building NextGen AI Infrastructure

[Nithin Stephen Raphael @RaphaelNithin]

@FCC @UnitreeRobotics https://t.co/4RlXLR5AZx

[Nithin Stephen Raphael @RaphaelNithin]

A new U.S. rule turns robot imports into a communications-security issue. @FCC is blocking new Chinese humanoid and quadruped models and can revisit existing authorizations. U.S. labs relying on lower-cost @UnitreeRobotics hardware may feel this before factories do.

[Nithin Stephen Raphael @RaphaelNithin]

@coursera @AndrewYNg https://t.co/Sdpg9xVVXm

[Nithin Stephen Raphael @RaphaelNithin]

A new education bet from @coursera puts about 13% of its March cash into @AndrewYNg's next company. Its $100M investment buys one-third of LearnVector, an AI tutoring startup for white-collar reskilling. That is a material balance-sheet bet, not a feature launch.

[Nithin Stephen Raphael @RaphaelNithin]

https://t.co/8O70x45sMs

[Nithin Stephen Raphael @RaphaelNithin]

AI detection has fresh capital and an old credibility problem. @pangram raised $9M, launched Pangram 4 and previewed an image detector. It claims 99%+ accuracy on mixed authorship; adoption will hinge on false positives outside its evaluations.

[Nithin Stephen Raphael @RaphaelNithin]

@AlphaDesignAI @Cadence @Synopsys https://t.co/wtCbD56o4v

[Nithin Stephen Raphael @RaphaelNithin]

Semiconductor verification is attracting agent money. @AlphaDesignAI added $60M to its Series A, bringing total funding to $131M. Its focus is finding functional bugs before tape-out, a costly part of chip design where @Cadence and @Synopsys are also adding agents.

[Nithin Stephen Raphael @RaphaelNithin]

@GeminiApp https://t.co/hBCQUyUCmA

[Nithin Stephen Raphael @RaphaelNithin]

A model can solve a language puzzle better than people and still fail to invent one. On novel Chinese xiehouyu, @GeminiApp's Gemini 3.1 Pro reached 92.6% accuracy, 24% above humans, while LLM-made riddles scored much worse. Reasoning and linguistic creativity are different tests.

[Nithin Stephen Raphael @RaphaelNithin]

https://t.co/GLxfcFMP9z

[Nithin Stephen Raphael @RaphaelNithin]

More explanation made developers trust AI code reviews more, but agree with them less. In a 34-person study, full explanations scored highest on perceived trust; shorter feedback reached 89.22% agreement. Good explanations may support scrutiny, not compliance.

[Nithin Stephen Raphael @RaphaelNithin]

https://t.co/B7OrQxIFfc

[Nithin Stephen Raphael @RaphaelNithin]

Shared agent memory needs a commit protocol. MemTX stages writes with evidence, permissions and provenance before they can drive actions, then repairs downstream records when beliefs are retracted. The authors report zero invariant violations across 5.5M tested states.

[Nithin Stephen Raphael @RaphaelNithin]

https://t.co/ZbWbWhisOe

[Nithin Stephen Raphael @RaphaelNithin]

An EEG foundation model can identify the dataset perfectly while struggling with the diagnosis. In a new stress test, dataset identity hit 1.000 AUROC, and classical features beat REVE on Korean dementia. Clinical AI needs stronger negative controls, not bigger embeddings.

[Nithin Stephen Raphael @RaphaelNithin]

@nvidia https://t.co/XxEq7hNC4O

[Nithin Stephen Raphael @RaphaelNithin]

A new Open Secure AI Alliance from @nvidia puts the open-vs-closed fight inside cybersecurity. Dozens of partners will build shared agent-security tools, while @nvidia contributes NOOA for testing, tracing and governing agent behavior. Security architecture is becoming policy.

[Nithin Stephen Raphael @RaphaelNithin]

@OpenAI https://t.co/oNVMXA6uFC

[Nithin Stephen Raphael @RaphaelNithin]

Job titles may be lagging the work already being done. @OpenAI says 43.5% of occupation-specific ChatGPT messages involve tasks associated with another occupation, based on 800,000+ U.S. work messages. AI adoption may show up first as fewer handoffs.